Extending the SISYPHUS III Experiment from a
Knowledge Engineering Task to a
Requirements Engineering Task

Debbie Richards, Tim Menzies

AI Department, School of Computer Science and Engineering,
The University of New South Wales, Kensington
{debbier, timm}@cse.unsw.edu.au

gzipped postscript version here

Abstract: The SISYPHUS III experiment draws attention to one of the major problems faced in knowledge engineering (KE), which is how to build systems based on multiple sources of expertise. In this circumstance, the KE task becomes a requirements engineering (RE) task. A problem with many RE approaches is that the cost of use is prohibitive, and therefore such approaches are rarely applied. We present an RE strategy designed to handle conflicting perspectives that is a small extension to current KE techniques. We implement this approach in the context of formal concept analysis (FCA) and ripple-down-rules (RDR) and describe an instantiation using the SISYPHUS III data. Our evaluation technique shows that the resolution operators have reduced the degree of conflict between viewpoints.

1. Introduction
2. Requirements Engineering: A Review
3. Our Framework
4. An Implementation
    4.1 Phase One: Requirements Acquisition using a Knowledge Based Approach
    4.2 Phase Two: Requirements Integration
    4.3 Phase Three: Concept Integration
    4.4 Phase Four: Concept Comparison and Conflict Detection
    4.5 Phase Five: Conflict Negotiation
    4.6 Evaluation
5. Requirements Engineering using SISYPHUS III
    5.1 Phase One: The Data and Knowledge Acquisition
    5.2 Phase Two: Requirements Integration
    5.3 Phase Three: Concept Integration
    5.4 Phase Four: Concept Comparison and Conflict Detection
    5.5 Phase Five: Conflict Negotiation
    5.6 Evaluation
6 Related Work
7 Future Work
8 Conclusion

1. Introduction

The SISYPHUS III experiment offers an excellent example of the similarity between the needs of knowledge engineering (KE) using multiple sources of expertise and those of requirements engineering (RE) from multiple perspectives. To a large extent in KE, the knowledge engineer decides which viewpoint to adopt. The RE approach, by way of contrast, is to negotiate a resolution between the owners of the different perspectives. However, is requirements engineering (RE) a complicated addition to current knowledge engineering (KE)? Will the process of rationalising multiple conflicting viewpoints slow down the production of an expert system? If the costs of RE are prohibitive, then RE will rarely be applied.

In this paper, we present a simple RE strategy that is an extension to current KE techniques. Given assertions in rulebases (the A-boxes) from different stakeholders, we generate and critique a concept hierarchy (the T-box). Conflicts recognised in the T-box can be used to drive negotiation strategies amongst the different stakeholders. A general framework for this approach is described together with an implementation using formal concept analysis (FCA) (Wille, 1982) and ripple-down rules (RDR) (Compton and Jansen, 1990).

A number of KBS have been developed using standard RDR KA techniques directly from the various SISYPHUS III sources. Each one of these KBS offer a quick and simple solution to the rock classification problem. We argue that by applying RE to these KBS we are able to develop a more representative and reliable KBS based on the combined viewpoints than the typical KE approach which leaves conflict resolution to the discretion of the KE who must become an expert in the domain to be able to make such decisions.. Modifications to KBs should optimise two criteria: competency and consensus. The focus of this paper is on consensus, not on the evaluation of the competency of the individual or combined rock classification KBS. Such an evaluation is problematic as precise evaluation standards have not been defined. Automatic methods of assessing competency are explored in Menzies and Compton (1997).

This paper is organised as follows. Section 2 introduces RE, Section 3 describes our RE framework and our implementation in found in Section 4. A case study is presented in Section 5 using seven knowledge bases built from the data for the SISYPHUS III (Shadbolt 1996) project. Our evaluation technique is described in Section 4.6 and is used in Section 5.6 to show how the resolution operators have halved the initial degree of conflict. Related work, future work and the conclusion are presented in Sections 6, 7 and 8, respectively.

2. Requirements Engineering: A Review

Requirements engineering is traditionally defined as: (e.g.) "the elicitation and formulation of requirements to produce a specification" (Easterbrook 1991, p.8). Current requirements engineering focuses on the maintenance of multiple concurrent viewpoints from different stake holders (e.g. Easterbrook or Finklestein et al 1994). Here, we offer a subtle re-expression of RE: "RE is the application of negotiation operators to reduce the differences between stake holder viewpoints". The advantage of this re-expression is that it can be mapped into an evaluation programme for RE, which we discuss more in 4.6 and demonstrate in 5.6.

There are a number of reasons why it is important to capture these multiple viewpoints rather than taking the approach that the different viewpoints must be captured into one specification (Easterbrook 1991, Ramesh and Dhar 1992). Tracking multiple perspectives is needed because:

Specification errors are often the cause of a poor choice between alternatives during the specification phase. By tracking multiple perspectives a history of the design rationale is provided so that when modifications are necessary they can be made more quickly and based on the background that formulated them in the first place.

Our concept of a viewpoint corresponds to Finkelstein et al's (1989) formalisation of a viewpoint which includes: a style, area of concern, a specification, a work plan and a work record. This allows for an individual to hold a number of viewpoints and removes some of the problems of equating a particular viewpoint with an individual. Each owner of a viewpoint we call a stakeholder. When managing different viewpoints, conflict between different stakeholders must be handled.

3. Our Framework

Our viewpoint management framework has the following five steps that are repeated until all stakeholders are satisfied.

4. An Implementation

Our general framework in Section 3 has not committed to any particular implementation choices. We continue now looking at this framework but within the context of our instantiation. As a result a number of restrictions on the generality of our implementation are imposed.

4.1 Phase One: Requirements Acquisition using a Knowledge Based Approach

We take a KBS approach to requirements acquisition. In this paper, we focus on the case where the inputs are multiple A-boxes provided from multiple experts, the T-boxes are empty and the set of cases for each A-box is not empty, that is A , T = and X , where A, T and X denote an assertional KBS, terminological KBS and set of cases, respectively. We place a further restriction that the A-box must be convertible into a decision table. It has been shown (Colomb 1993) that any decision tree or propositional KBS may be converted into a decision table. Other work (Richards and Compton 1997b) has shown how ripple-down rule (RDR) systems can easily be converted into decision tables. Conversion to a decision table is also suitable for production rule-based systems and has been applied to a number of CLIPS KBS. For our purposes we distinguish between RDR systems and propositional rulebases. We refer to the latter as "standard rules".

This phase is also the maintenance phase for once one cycle is completed it is vital to ensure that the changes made to the explanation system (T-box) output from Phase Five are reflected in the appropriate individual and shared performance systems (A-boxes). This study proposes the use of multiple classification RDR (MCRDR) (Kang, Compton and Preston 1995) for knowledge acquisition (KA) and knowledge representation (KR). We adopt RDR because maintenance in RDR is a simple task that can be performed by the user and does not suffer from the side-effect problem which occurs in typical rule-based systems (Soloway, Bachant and Jensen 1987). KA using RDR has been designed to be performed by the user and as Herlea (1996) points out user involvement is a critical issue in making RE work. An additional benefit of RDR, as mentioned above, is that the rule pathways map directly into a decision table and do not need intermediate conclusions to be mapped to primitive conditions as many rule bases require.


Figure 1. An MCRDR KBS.

The highlighted boxes represent rules that are satisfied for the case {a,d,g,h,k}. We can see that there are two independent conclusions for this case, Class 2 (Rule 2) and Class 5 (Rule 10). Rule 5 had been the cause of a conflict between viewpoints. To resolve this conflict it was decided that attribute g should be dropped. As described in Figure 2, the STOPPING RULE is used to say that this pathway should not fire, so even though the case satisfies Rule 5 that rule is stopped from being reported. We can see that Rule 10 now replaces the rule pathway for Rule 5 dropping the attribute g.

The RDR approach to KA is to run a case and show the user the system-assigned conclusion/s. If the user agrees with the conclusions given then they process the next case. If they do not agree with a conclusion they take the option to reclassify the case. Reclassification involves specifying the correct conclusion and picking some features in the case that justify the new conclusion. These features form the conditions of the new rule. The new rule is added as an exception to the rule that gave the misclassification. The case that prompts a rule to be added is stored in association with the new rule. When a new rule is added, the rule must distinguish between the present case and all the stored cases that can reach that rule. To do this, the expert is required firstly to construct a rule which distinguishes between the new case and one of the stored cases. If other stored cases satisfy the rule, further conditions are required to be added to exclude a further cases and so on until no stored cases satisfy the rule. Stopping rules, which prevent an incorrect conclusion by providing a null conclusion are added in the same way. Surprisingly the expert provides a sufficiently precise rule after two or three cases have been seen (Kang, Compton and Preston 1995). Multiple Classification (MCRDR) is defined as the triple <rule,C,S>, where C are the children/exception rules and S are the siblings. All siblings at the first level are evaluated and if true the list of children are evaluated until all children from true parents have been exhausted. The last true rule on each pathway forms the conclusion for the case. Figure 1 shows an example of an MCRDR.

The greatest success for RDR has been the Pathology Expert Interpretative Reporting System (PEIRS) (Edwards et al 1993). PEIRS went into routine use with approximately 200 rules and grew in a four year period (1990-1994) to over 2000 rules. The system was maintained by the expert and the 2000 rules represent a development time of approximately 100 hours.

As noted in our general framework, the input to Phase One also includes a set of cases. The importance of the set of cases is twofold. Firstly, on a general level, the cases are used in the negotiations as counterexamples for discussion. Secondly, using the RDR approach as described above, or other case-based technique, the cases are also used for initial KA and for modification of other views. The way this works is that when a concept is found to be in conflict, as one of our modification strategies outlined below (see Figure 2), we pass the case or cases associated with that concept to the other stakeholder for KA. This should either resolve the conflict or at least ensure that both parties have given their views given the same set of criteria. In Related Work we suggest how we can obtain cases.

4.2 Phase Two: Requirements Integration

Requirements integration is the process of ensuring that all viewpoints are in formats that can be compared. Adopting our general framework, it may be that viewpoints have been captured using different KR. To avoid the requirement of mapping from all KR's used into all other KR's, that is N2 mapping schemes, we convert all KRs into one format so that we only need 2N mapping schemes. As mentioned in Phase One, in our current implementation we can use any representation that maps into a decision table so that a common approach to subsequent phases can be taken. It will become apparent in the next section why we have the decision table format restriction.

4.3 Phase Three: Concept Integration
In our general framework, an explanation system could already exist (that is, T in Phase One). Alternatively, it could be supplemented or built in this phase. In the current work, we restrict ourselves to the case where T = . The approach we have chosen is to begin with a performance system and later derive the explanation system. We start with a set of privately owned and defended A-boxes (A1..Ai) written by some experts (V1..Vi). The knowledge base also includes some data structures generated from previous cycles of our five phases. These structures are:

For more on these tables, see Section 4.5.

We have taken the approach of starting with a performance system (A-box) and using that to derive an explanation system (T-box) because we see that defining and building models is inherently difficult and flawed. Easterbrook (1991) points out that one of the reasons why systems fail to meet the user's needs is that the original mental model of the user has not been captured in the final design model. We see the difficulty in capturing mental models as a contributing factor to the bottleneck associated with KA. We now explain how we use FCA to make the leap from A-box to T-box.

A concept in FCA is comprised of a set of objects and the set of attributes associated with those objects. The set of objects forms the extent of the concept while the set of attributes forms the intent of the concept. Knowledge is seen as applying in a context and can be formally defined as a crosstable. We interpret the decision table in Figure 5 as a formal context where the rows are objects and the columns are attributes. An X indicates that a particular object has the corresponding attribute. This crosstable is used to find formal concepts.

In Figure 5 we have the formal context "First four MCRDR rules for Card Sort 5" with the set of objects = {1-%NC000, 2-%AD000, 3-%AN000, 4-%BA000}, where the object name is made up of the rule number and the conclusion code, and set of attributes = {1=1, grain_size = coarsely_crystalline, silica = very_high, colour = light, grain_size = fine, silica = intermediate, silica = basic }. Note that the condition 1=1 is used for the default conclusion %NC000. The crosses show where a relation between the object and attribute exists, thus the set of relations = {(1-%NC000, 1 = 1), (2-%AD000, grain_size = coarsely_crystalline) (2-%AD000, Silica=very_high),, (4-%BA000, silica=basic)}. Each row in the crosstable represents a concept. By finding the intersections of sets of attributes and the set of objects that share those attributes we are able to form new higher level abstractions. The set of concepts can be ordered using the subsumption relation on the set of all concepts which can be used to form a complete lattice. For a more detailed and formal treatment of our approach see Richards and Compton (1997b).

In Figure 6 the concepts are shown as small circles and the sub/superconcept relations as lines. Each concept has various attributes and objects associated with it. The full labelling for each concept has been reduced for clarity by removing all attributes from a concept that can be reached by ascending paths and removing all objects from a concept that can be reached by descending paths. Thus, the concept lattice provides "hierarchical conceptual clustering of the objects (via the extents) . and a representation of all implications between the attributes (via its intents)" (Wille 1992, p.497).

4.4 Phase Four: Concept Comparison and Conflict Detection

A number of researchers offer different sets of conflict types (e.g. Easterbrook 1991 and Schwanke and Kaiser 1988). In this paper, we use the four quadrant model of comparison between experts developed by Gaines and Shaw (1989). This model classifies two conceptual models as being in one of four states:

Consensus is the situation where experts describe the same concepts using the same terminology.

Correspondence occurs where experts describe the same concepts but use different terminology.
Conflict is where different concepts are being described but the same terms are used.
Contrast is where there is no similarity between concepts or the terminology used.

In this paper we generally take a broader view of conflict to encompass inconsistencies that include the states of contrast, correspondence and conflict. Gaines and Shaw's model, however, does offer us greater precision in describing the nature of the conflict which is important in deciding how it can be handled. We more formally define the states of consensus and contrast according the FCA notion of a concept as a related set of attributes and objects. For an example of each of the four states as they may appear on a concept lattice see section 5.4.

Consensus exists where all of the attributes and objects of a concept in one viewpoint match with a concept in another viewpoint. The search for a match is sequential and terminates when a match is found or when all the concepts in the other viewpoint have been compared. Discovering the consensus between conceptual models is important for establishing common grounds from which differences can be viewed. There is no point considering differences if there is no similarity in the first place. Contrast can be viewed as the opposite to consensus and exists when none of the attributes or objects in a concept in one viewpoint can be found in any of the concepts in another viewpoint. The search for a contrasting concept terminates when a match on any attribute or object is found or when all concepts in the other viewpoint have been processed.

A concept not in a state of consensus (match found in another viewpoint) or contrast (completely different to all concepts in another viewpoint) is then either in a state of correspondence or conflict. The key to deciding which state it belongs to depends on the terminology. In our approach it would be up to the stakeholder to decide whether the terminology used for an attribute or object was the cause for two concepts not appearing at the same node. If more assistance for the user is desired, Gaines and Shaw (1989) have shown that the repertory grid technique, which our crosstable could be mapped into, can be used to identify where terminology is the cause of inconsistency. Having detected various conflicts we now consider how to resolve them.

4.5 Conflict Negotiation

Before we can decide how to fix a detected inconsistency we need to provide a conflict resolution strategy. There are a number of resolution methods which include negotiation, arbitration, coercion and education (Strauss 1978). Negotiation is the most appropriate within the assumed context of parties of equal status and ability. As Easterbrook (1989) points out, a good solution will require creativity and creativity is not something that can be automated. However, since automation is a fundamental goal of requirements engineering research we extend our approach beyond a general, genial chat by offering as much automated assistance for this step as possible.

Each RE researcher appears to use a different set of resolution strategies (e.g. Easterbrook 1991, Thomas 1976). Easterbrook and Nuseibeh (1996) offer five categories that cover the actions we have found necessary. These are:

Resolving conflict will involve performing modifications. If the cause of disagreement is differences in terminology (correspondence in the Gaines and Shaw four state model) then one technique is to up-date all views to conform to an agreed upon set of terminology. This option is probably not satisfactory to the various stakeholders and also means that the history of changes is being lost or altered. A simple and more appropriate solution is to use a subsumes table which maps terms from individual viewpoints into a shared terminology which are then used for comparison.

Another way in which conflict may be resolved is through the addition or deletion of attributes or objects. The conflict may be that the set of attributes or objects are partially shared by another concept. To bring these concepts into a state of consensus it may be decided to drop or add attributes or objects. As mentioned in Phase One, part of our automated support for negotiation is the ability to produce a case associated with the object (rule) that is in question. The cases associated with all objects that can be reached by downward paths are also relevant to the discussion. The closer, distance measured in number of objects separating the two nodes, the object is to the concept in question the more relevant the case should be considered. New attributes or objects could also be added by showing the associated case to the other party and using that case for KA. Alternatively, if a hypothetical case is shown to be impossible, then the rules based on this case should be dropped. The decision of what action to take is made in this last phase and performed in Phase One. In Figures 2(a) and (b) we provide a summary of the strategies applicable to standard rules and MCRDR. Note that:

Add Attribute Delete Attribute Add Object Delete Object
Standard Rules Add attribute to rule. Check effect on other rules. Remove attribute from rule. Check effect on other rules. Add new rule. Check effect on other rules. Remove rule. Check effect on other rules.
MCRDR Defined - use existing KA approach Add stopping rule to rule in error. Add new rule at top level with the old rule minus the attribute to be removed. Perform KA using the case associated with the concept which has the desired object. Add a stopping rule.

Figure 2(a) The resolution strategy for handling attribute and objects.

Adding Real Cases Deleting Real Cases Adding Hypo-thetical Cases Deleting Hypothetical Cases
Standard Rules N/A N/A N/A N/A
MCRDR Show to all ILLEGAL Show to all Drop related rules - check refinement rules if they should be dropped or stopped and a new rule, without the dropped rule conditions, added.

Figure 2(b) The resolution strategy for handling cases.

We plan to strengthen our negotiation strategies by offering filtering rules which guide the dialogue between the stakeholder and the system. One filter is the use of preference criteria to guide the stakeholders in deciding what part of the view should be considered first as a candidate for change. If we consider the concept lattice structure where objects belonging to a concept are reached by descending paths and attributes are reached by ascending paths then we can say that if an attribute at the bottom of a pathway or an object at the top of a pathway is to be removed then we know that no other concepts will be affected and this can be performed without further investigation. This strategy can be useful for example, if it had been agreed upon that only one of two attributes were necessary and one was at the bottom of a pathway and the other higher up, then it would be advisable to remove the lower attribute. Other templates or KA scripts (Gil and Tallis 1997) can be used to guide the user with revising their KBS and ensuring that each of the rules related to the change are modified and tested.

The last four resolution strategies are relevant for situations in which a complete resolution can not be negotiated and each one has its appropriate usage. For example, ignoring is a useful strategy where the issue is not that important or pursuing it is not worth the effort or harm it may cause to the end solution. These approaches can be termed as living with inconsistency or 'lazy' consistency (Narayanasway and Goldman 1992) and can be compared to fault-tolerant systems that continue to function after non-critical failures occur. It has been argued that enforcing removal of all inconsistency "constrains the specification unnecessarily" and "tends to restrict the development process and stifle novelty and invention" (Finkelstein et al 1994, p.2 & 4). They see that consistency is necessary within a viewpoint but partial consistency between viewpoints is allowable.

We also accept that living with inconsistency will be necessary and use tags to identify the status of the conflict. The use of tags is similar to the use of "pollution markers" (Balzer 1991) that act as a warning that code may be unstable or that the users should carefully check the output. Pollution markers can be used to screen inconsistent data from critical paths that must have completely consistent input. If it is the concept that is being circumvented, ignored or delayed, we mark the concept in the shared T-box since there is not necessarily a one-to-one correspondence between rules and concepts. This updated T-box is used as input in the next T-box generation. When it comes to rules we only offer the strategy of circumvention to support avoidance of certain unstable parts of the requirements by tagging a rule as "circumvented" in the individual A-boxes. When the new T-boxes are generated these rules will not be included.

The resolution strategies shown in Figures 2(a) and (b) are also applicable to the strategy of amelioration. However, the result is not consensus but a reduction in the extent of the conflict. Amelioration results in bringing concepts closer together. If we think in terms of the concept lattice we would be shortening the distance between the two concepts.

4.6 Evaluation

To evaluate that our RE operators are reducing the conflict between viewpoints we need two sub-routines: ·

These RE sub-routines can now be evaluated as follows. Different negotiation operators can be assessed via their observed effects on the distance metric: good RE operators have a big impact on this distance. Operators also have a construction cost: good operators should be easy to build. An ideal RE operator is cheap to build, cheap to use, and reduces the distance metric considerably (though, in practice, some trade-offs may be required between operator complexity and operator effectiveness).

We use our negotiation operators as already defined and compare the degree of conflict before and after by computing a distance metric for each concept in each viewpoint compared to each other viewpoint and taking the total of these scores. We assign of score of 0 to a concept found to be in a state of consensus with a concept in another viewpoint, since the distance between them is zero. For concepts in a state of conflict we take the number of attributes (conditions) that they have but do not share divided by the total number of attributes. This assumes that the two concepts share the same object (conclusion). If they do not then it appears that they are not meant to represent the same concept so that comparison is not meaningful. For concepts in a state of contrast (no partial or complete match in the other viewpoint) we assign a score of 1, which is the same result as if we used the conflict measure since the number of attributes not shared divided by the number of attributes is equal to one. Concepts in a state of correspondence are treated the same as concepts in conflict since we are ignoring the reason for the differences and are just interested in the size of the difference. Once terminology differences are reconciled such concepts will move into one of the other states and be handled accordingly.

We have used these measures in Section 5.6 to determine how well our RE strategy is working. Since in the next section we restrict ourselves to looking at one conclusion at a time in all viewpoints the measures described are appropriate. Of greater difficulty is determining the degree of conflict if we do not know which concept in the other view represents a similar concept. We either need to get the experts to label their concepts so that, for example, two concepts labelled "cat" would be compared or we can compare the concept in one viewpoint with all others in the other viewpoint finding the one that shares the largest set of attributes on the assumption that it is the concept which most likely represents the same concept. For the purposes of the example in the next section we did not need to resolve this problem and further work on distance measures is part of our future work.

5. Requirements Engineering using SISYPHUS III

The problem statement and scope of SISYPHUS III is given as (Shadbolt 1996):


We extend the problem description a little further for the purposes of this case study. The KA material supplied with the first phase includes: 5 card sorts, 4 laddered grids, 5 structured interviews, 4 self reports and 4 repertory grids. The card sorts and repertory grids provide useful frameworks that require the least amount of data extraction compared to the other techniques that use natural language. Thus, we not only want to develop a system that classifies a set of igneous rocks correctly and offers an explanation of the conclusion but we want to use data from multiple and conflicting sources of expertise to build this system. These multiple sources of expertise impose a further requirement on the project not indicated in the problem statement above which is to develop a strategy for handling any differences. Thus, the SISYPHUS III data provides an ideal source of different models covering the same domain in which to test out our RE framework.

The RDR submission produced a number of KBS from a combination of these sources. The systems were built by a geological and novice KE in the same manner that is described in section 5.1. Differences in terminology, measurement and inconsistencies between the sources were reconciled using the rule of thumb "choose the option that most experts agreed on", where each KA source was treated as an expert. This is a reasonable heuristic but one that can not always be applied since in some cases it requires interpretation to decide that two concepts match. Thus, to some extent the final system is a product of what our novice KE thought was right. It seemed reasonable that the approach outlined in this paper for reconciling conflict in multiple viewpoints would provide a better way of arriving at a final though not complete model that took into account each perspective without the bias of the KE as the method of resolution.

In the next subsections we work through the five phases already presented showing how the proposed framework can be implemented. Note that the process is implemented on a small scale - that is the conclusions are compared one at a time rather than comparing all the complete viewpoints in one go. The main reason for this is the more information we are dealing with, in particular representing visually, the more incomprehensible the diagram becomes. While this approach may be seen as time consuming it is not necessarily the most inefficient because problems can often be better identified and resolved when broken down into smaller problems. One of the benefits is that as we begin to resolve conflicts, for example reconcile differences in terminology through the subsumes table or tag concepts that should be ignored, there will be less errors or differences as we progress.

5.1 Phase One: The Data and Knowledge Acquisition

To compare the models represented in each of the data sources mentioned above, we had our novice KE develop individual KBS using MCRDR for each of the card sorts and laddered grids. For the purposes of this case study we did not use the other material even though supporting or extra information had originally been gleaned from them and added to the final RDR KBS solution. We did this because we wanted the models to represent only the data directly specified as related to them and to avoid inclusion of the KE's own interpretation. The terminology used in each KBS is the terminology provided in the source KA material. The card sorts contained differences in the number of attributes used, the attributes chosen, the values assigned to attributes and the terminology of attributes and objects. It was these types of differences we wished to reconcile. One alternative is to force a common set of terminology and scales to be used. However, it is important to stakeholders/experts that they be allowed to express their models in their own words. This is supported by the work on personal construct psychology, where subjects preferred to use their own constructs over ones supplied to them (Shaw 1980).


#1 10-03-1997 09:16:57
GRAIN_SIZE           COARSELY_CRYSTALLINE
SILICA                     VERY_HIGH
COLOUR                  LIGHT
OLIVINE                   UNLIKELY
MICA                        LIKELY
#2 10-03-1997 09:16:57
GRAIN_SIZE           FINE
SILICA                      INTERMEDIATE
COLOUR                   FAIRLY_LIGHT
OLIVINE                   UNLIKELY
MICA                        UNLIKELY

#3 10-03-1997 09:16:57
GRAIN_SIZE           FINE
SILICA                     BASIC
COLOUR                  DARK
OLIVINE                   LIKELY
MICA                        UNLIKELY


Figure 3: Extract from Card Sort 5 Case File.

The data in the card sorts were used directly to develop cases. In Figure 3 we can see the first three records in the Card Sort 5 case file. For the restricted study performed in this paper, a case file was developed only for Card Sorts 1, 2, 3 and 5 and Laddered Grids 1 and 3, and will be referred to as C1, C2, C3, C5, L1 and L2, respectively. However, for the purpose of the example using Adamellite a partial KB for Card Sort 4, C4, which covered that rock was developed. Using the appropriate case file, KBS were developed by running an inference on each case. If the conclusion given by the system was incorrect, a new conclusion was assigned and the attribute value pairs that justified that conclusion were selected as the rule conditions. The conditions were selected based on intuition and the difference list that shows the differences between the current case and the case that gave the misclassification. The MCRDR rule file developed from the Card5 case is shown in Figure 4. For the rest of this example, each KBS is treated as a different expert viewpoint.


0
11
12
1 0 11 0 %NC000 : 1 = 1

2 1 0 0 %AD000 : (GRAIN_SIZE = COARSELY_CRYSTALLINE) & (SILICA = VERY_HIGH) & (COLOUR = LIGHT)
3 1 0 2 %AN000 : (GRAIN_SIZE = FINE) & (SILICA = INTERMEDIATE)
4 1 0 3 %BA000 : (GRAIN_SIZE = FINE) & (SILICA = BASIC)
5 1 0 4 %DI000 (GRAIN_SIZE=COARSELY_CRYSTALLINE)&(SILICA=INTERMEDIATE)&(COLOUR=FAIRLY_LIGHT

6 1 0 5 %DO000 : (GRAIN_SIZE = NOT_COARSE) & (SILICA = BASIC)
7 1 0 6 %DU000 : (GRAIN_SIZE = COARSELY_CRYSTALLINE) & (SILICA = ULTRABASIC) & (COLOUR = DARK)
8 1 0 7 %GA000 : (GRAIN_SIZE = COARSELY_CRYSTALLINE) & (SILICA = BASIC) & (COLOUR = DARK)
9 1 0 8 %KE000 : (GRAIN_SIZE = ?)
10 1 0 9 %GR002 : (GRAIN_SIZE = NOT_COARSE) & (SILICA = VERY_HIGH)
11 1 0 10 %RH000 : (GRAIN_SIZE = FINE) & (COLOUR = LIGHT)

Figure 4: The rule file for Card Sort 5.

The first rule number, the total number of rules and the number of cases associated with the rules are given in the first three lines. The default rule is on line four. The structure of each rule is rule number, parent rule number, child rule number, sibling rule number, conclusion code and the conjunction of conditions after the colon.

5.2 Phase Two: Requirements Integration

As described in our framework, each KBS is converted into a common format, specifically a crosstable or decision table. Figure 5 shows the crosstable, which is a formal context, for the first four MCRDR rules for Card Sort 5.


Figure 5: Crosstable of the first four MCRDR rules for Card Sort 5

1=1 grain_size= coarsely_crystalline silica= very_high colour= light grain_size= fine silica= intermediate silica= basic
1-%NC000 X
2-%AD000 X X X X
3-%AN000 X X X
4-%BA000 X X X

5.3 Phase Three: Concept Integration

In this phase the crosstable is used to derive the ordered set of formal concepts which can be shown as a concept lattice. When dealing with multiple KBS there are two main alternative approaches to implementation of this phase. The crosstables or formal contexts output from Phase Two can be combined into one formal context with the source of each row annotated for identification OR the concepts for each formal context can be derived and then compared. The simplest method with the existing MCRDR/FCA tool is to combine all the individual formal contexts into one formal context and perform queries or comparisons on them all. However, handling such a large context has its problems. As previously mentioned:

"the diagrams for activities of any reasonable complexity become very difficult to visualise and understand" (Gaines and Shaw 1993, p. 59).

To address this problem the MCRDR/FCA system provides a selection screen with 13 combinable options that allow the user to narrow their focus of attention to selected part/s of interest in the knowledge base. This approach may not be totally satisfactory as it requires a methodical approach to comparing the various parts of each KBS and may miss some conflict that is not being picked up because the relevant selection criteria was missed. However, if our major goal is correct classification of a particular rock then it seems reasonable and manageable to explore each conclusion individually from the combined viewpoints to see what conflict exists. Figure 6 shows the concept lattice for the conclusion %AD000- Adamellite for the seven viewpoints which are annotated as C1, C2, C3, C4, C5, L1 and L3..

Figure 6: The Concept Lattice for the Conclusion %AD000- Adamellite based on seven KBS.

5.4 Phase Four: Concept Comparison and Conflict Detection

The concept lattice in figure 6 may be analysed and the nature of the differences between concepts determined. The C5 viewpoint for Adamellite is in contrast since none of its attributes are shared with any of the other viewpoints. However, some of these differences appear to be terminology related. There is consensus between C1, C2, C3 and L2 that GRAIN_SIZE=COARSE but in C5 the GRAIN_SIZE=COARSELY_CRYSTALLINE. This appears to be a correspondence type of conflict because of differences in terminology. There appear to be other correspondence errors. The attribute QUARTZ is used in C1, C2 and C4 with the values ALWAYS, SATURATED and 10%, respectively. The value of OLIVINE in L1 and L3 is NO and in C2 the value is ABSENT. In C5 the COLOUR=LIGHT and in C1 the COLOUR=LEUCRATIC. The dictionary meaning of "leuco" is white (Macquarie Dictionary), so it appears that the terms in these two concepts are compatible. It also appears that DARK_MINERAL=LT30 also indicates a lightness of colour. The differences in the terminology used for the values assigned to GRAIN_SIZE, QUARTZ, OLIVINE and COLOUR can be reconciled by using the subsumes table to map to a common term as shown in Figure 7.

The value assigned to OLIVINE in C1 is ALWAYS and represents a conflict where the terms are compatible but the concept is obviously the opposite to the concepts in L1, L2 and C2. There is consensus between L1 and L2 that PLAGIOCLASE = NO but these concepts conflict with the concepts FELDSPAR=1_3TO2_3PLAGIOCLASE for C3 and FELDSPAR>2/3ORTHO&<1/3PLAGIOCLASE for C4. These various conflicts need to be resolved which takes us to our next stage.

5.5 Conflict Negotiation

Original Term Subsumed Term
GRAIN_SIZE=COARSLY_CRYSTALLINE GRAIN_SIZE=COARSE
QUARTZ=SATURATED QUARTZ=ALWAYS
QUARTZ=10% QUARTZ=ALWAYS
OLIVINE=ABSENT OLIVINE=NO
COLOUR=LEUCRATIC COLOUR=LIGHT
FELSIC_COLOUR=

WHITE&LARGE_CRYTALS

COLOUR=LIGHT
DARK_MINERAL=LT30 COLOUR=LIGHT

Figure 7: The Subsumes Table

Reconciliation of conflicts in terminology is a good place to start in the negotiation process. Those terms identified to be equivalent are put in the subsumes table, see Figure 7. We update this table and go through phases 1-5 again. In Figure 8 we can see the effect of the subsumes table on the original conflict in Figure 6.

Figure 8: The updated Line Diagram after the Subsumes Table in Figure 7 has been included.

In addition to terminological differences we have some conceptual differences. It is a limitation of this study that no experts were available to perform this actual reconciliation as would be expected in a real RE problem. Even if we could have found a number of geological experts none of them would have been the owners of the various viewpoints in the card sorts or laddered grids so that would not have overcome this limitation. It was therefore necessary to simulate how some of these differences can be resolved with some hypothetical experts. The following scenario is offered:

Figure 9: Users can flag a concept to be "ignored" on the CMatrix Screen by clicking on the Concept No and entering the Tag Code into the pop-up InputBox.


All of the changes mentioned above are reflected in our final diagram in Figure 10. Note that even though the number of concepts has only reduced by 1 the concepts are much less complex. In figure 6 GRAIN_SIZE=COARSE offered the most, but not total, point of agreement. Now all views agree with this and there are more attributes shared by viewpoints that previously only appeared in one viewpoint. As shown visually in Figure 10, the viewpoints in Card Sorts 1,2,3, and 5 are more similar to each other than the viewpoints in Laddered Grids 1 and 3 which are similar to each other. For those finding the diagram unfamiliar each of the viewpoints can be summarised as:

C1- GRAIN_SIZE=COARSE;COLOUR=LIGHT;OLIVINE=NO;QUARTZ=ALWAYS
C2- GRAIN_SIZE=COARSE; OLIVINE=NO;QUARTZ=ALWAYS
C3- GRAIN_SIZE=COARSE;COLOUR=LIGHT
C4 - GRAIN_SIZE=COARSE;COLOUR=LIGHT;QUARTZ=ALWAYS
C5- GRAIN_SIZE=COARSE;COLOUR=LIGHT;SILICA=VERY_HIGH(DELAYED)
L1- GRAIN_SIZE=COARSE;OLIVINE=NO;PLAGIOCLASE=NO;PYROXENE=NO(DELAYED)
L3- GRAIN_SIZE=COARSE;OLIVINE=NO;PLAGIOCLASE=NO

Figure 10: The final Line Diagram screen from this round of negotiations. Some attributes have been dropped or added to views, concepts have been tagged to be circumvented (not shown) or delayed (shown). The case for concept 8 is shown to assist with negotiation. There is considerably less conflict now than in Figure 6.

As a further test on the amount of agreement in our T-boxes we can load all six A-boxes, recall only six were complete, and run cases to see the amount of agreement. In Figure 11 we tested our case for Adamellite and found that C1(KB1), C2(KB2), C3(KB3), C5(KB4) and L3(KB6) all conclude %AD000. The rule for Adamellite in L1 does not fire because it has the condition (PYROXENE=NO) which is not present in the case. If the value for PYROXENE in the case were modified to NO instead of missing then all KBS would correctly classify this rock specimen. Note that when this case was originally run before the modifications none of the KBS gave the conclusion %AD000. The reader is invited to verify this by using the case in Figure 11 against the concept lattice in Figure 6 treating the attributes that can be reached by an ascending pathway from a conclusion as the rule conditions.

Fig 11: Running a Case against Multiple KBS as a Test on the Degree of Agreement


5.6 Evaluation

Using the measures described in 4.6 we computed the degree of conflict at three points, at the beginning, after updating the subsumes table and at the end. When the process described is performed for the all rock classification we could have computed the reduction in conflict after each operator was applied to determine which operators were the most effective. The example offered here is not sufficient data for such a comparison. The only concepts in consensus are concepts matched against themselves, there are a number of contrasting concepts shown by the cell entries which equal 1, such as C1-C5 and C1-L1. Concepts shown as a fractions are in conflict where the numerator is the number of attributes not shared and the denominator is the total number of attributes in both concepts. The total degree of conflict for a viewpoint is shown as a decimal for better comparison.

C1 C2 C3 C4 C5 L1 L3 TOTAL
C1 0 5/7 5/7 6/8 1 1 5/7 4.89
C2 5/7 0 4/6 5/7 1 1 4/6 4.76
C3 5/7 4/6 0 5/7 1 1 4/6 4.76
C4 6/8 5/7 5/7 0 1 1 5/7 4.89
C5 1 1 1 1 0 1 1 6.00
L1 1 1 1 1 1 0 2/6 5.33
L3 5/7 4/6 4/6 5/7 1 2/6 0 4.10
34.74

Table 1: The degree of conflict between each viewpoint for the %AD000-Adamellite conclusion before RE.

In Table 1 we can see that all viewpoints are in conflict with others, with views C1, C2, C3, C4 and L3 having similar degrees of conflict. From this table we can see that viewpoint C5 is in complete contrast with all other views, thus it has the highest degree of conflict, followed by L1 that only shares some attributes with L3. In Table 2, we can see that the total degree of conflict for each viewpoint is lower after similar terms have been reconciled using the subsumes table, even though it has not reduced the conflict in all cells. The total degree of conflict for all viewpoints has reduced from 34.74 before we began our resolution strategies to 24.88 after we applied our first strategy of reconciling terms. It is interesting to note that all views except L1 now have similar, though lower than before, degrees of conflict. This shows that much of the conflict originally in C5 was due to differences in terminology, which we have already discovered in our previous discussions. Very little of the conflict in L1 appears to be terminology related and as is shown in Table 3 after we have completed this round of negotiations, L1 remains the viewpoint most in conflict with the others.

C1 C2 C3 C4 C5 L1 L3 TOTAL
C1 0 3/7 3/7 2/8 1 1 5/7 3.25
C2 3/7 0 4/6 3/7 3/6 4/6 2/6 3.02
C3 3/7 4/6 0 3/7 2/6 1 4/6 3.52
C4 2/8 3/7 3/7 0 3/7 1 5/7 3.25
C5 1 3/6 2/6 3/7 0 1 4/6 3.35
L1 1 4/6 1 1 1 0 2/6 5.00
L3 5/7 2/6 4/6 5/7 4/61 2/6 0 3.48
24.88

Table 2: The degree of conflict after mapping similar terms using the Subsumes Table.

After we have applied the resolution strategies in Figure 2 and used our various resolution Tag Codes to produce the final concept lattice in Figure 10, the degree of conflict remaining is shown in Table 3. As noted before we have not removed all conflict but it has reduced by 53% from 34.74 to 18.49. We conclude that our resolution operators have reduced the degree of conflict.

C1 C2 C3 C4 C5 L1 L3 TOTAL
C1 0 1/7 2/6 1/7 3/7 4/8 3/7 1.98
C2 1/7 0 3/5 2/6 4/6 3/7 2/6 2.50
C3 2/6 3/5 0 1/5 1/5 4/6 3/5 2.60
C4 1/7 2/6 1/5 0 2/6 5/7 4/6 2.39
C5 3/7 4/6 1/5 2/6 0 5/7 4/6 3.01
L1 4/8 3/7 4/6 5/7 5/7 0 1/7 3.17
L3 3/7 2/6 3/5 4/6 4/6 1/7 0 2.84
18.49

Table 3: The degree of conflict between each viewpoint for the %AD000-Adamellite conclusion after RE.

6. Related Work

Starting with a performance system and deriving an explanation system, is in complete contrast to mainstream KA research where the focus is on building problem-solving (Chandrasekaran and Johnson 1993, McDermott 1988, Puerta et al 1992, Schreiber, Weilinga and Breuker 1993 and Steels 1993) and/or ontological (Guha and Lenat 1990, Patil et al 1991, Pirlein and Studer 1994) models first and using these to develop a performance system. As explained earlier, conceptual models are difficult to capture and unreliable and we prefer to capture simpler models based on the performance knowledge that can be demonstrated and observed. RDR has given us a reliable method for capturing and maintaining performance knowledge and FCA is the mechanism that lets us derive the explanation system.

In our framework, cases play a critical role and we have assumed that cases are available. In our use of the SISYPHUS III data we have had a source of cases but this may not be always be so, particularly when dealing with software specification requirements. One viable option for the purposes of RE is use cases since they satisfy our need for sets of attributes and outcomes and are "primarily an approach to discovering requirements from a user-centred viewpoint" (Rumbaugh 1994, p.23).The first step is to enumerate the actors, which are external agents that require services from the system, and then the use cases. Specific values, not generalisations should be plugged into the cases so that thinking is grounded in precise examples. Rumbaugh suggests first building a system which contains the domain model and then the application model using use cases. We can equate the domain model and application model to our T-box and A-box, respectively. What we are advocating is to use the cases to build the A-box, the rules, from which we derive the T-box, the concept hierarchy. The use-cases could be used in conjunction with RDR or as direct input into the decision table and maintenance would consist of modification to the use cases.


Prior work in this area has been presented as verification or maintenance research (Compton et al 1992, Richards and Compton 1997a). We here claim that an RE technique could be viewed as a technique for knowledge maintenance (KM) and verification and validation (V&V). Menzies (1997) argues that knowledge management techniques amount to a small number of activities processing less than a dozen types of knowledge. At different points of the software lifecycle, the emphasis is on certain activities processing some of the knowledge types:

The only difference between MCRDR/FCA for requirements engineering and MCRDR/FCA for maintenance is that we can use one more revision assessment operators in the maintenance case (regression analysis over behavioural knowledge). We speculate that this work is the first step towards a succinct toolkit to support knowledge engineering and software engineering activities right across the life cycle.

7. Future Work

We have defined here a preliminary exploration of our framework and shown parts of the toolkit in operation. Clearly, the next stage is full implementation of this approach. Such an implementation would have to handle a variety of interface details (e.g. appropriate displays of large concept lattices). Such an implementation could be an extension of Richard's existing MCDRD/FCA tool (seen in Figures 6 and 8).

Our approach aims to minimise the conceptual distance between A-boxes from different experts. Our decision to accept that some conflict will always remain means that we need an evaluation and termination strategy that lets us test whether our framework is reducing the conflict and when enough conflict has been resolved to allow the specification to be used. As demonstrated in Sections 4.6 and 5.6 we want to track over-time whether which each subsequent cycle we are moving towards consensus and these statistics should be computed when the resolution strategies for the current cycle have been applied. However, the statistics we used were easily computed because we knew which concepts to use for comparison. As pointed out in Section 4.6, it we want to compute an overall degree of conflict within each KBS it may be harder to determine what concepts should be compared and how these scores should be used. Currently, we are working on a graph-theoretic distance measure between T-box lattices.

8. Conclusion

We have argued here for a novel view of RE. Standard RE is an early-software lifecycle issue. The viewpoint resolution technique discussed here can be performed whenever we have some A-box (rules) and cases. Initially, cases will be hypothetical and the rules sets small (snippets of known business processes). Finally, cases will be "live" data and rule sets will be large. In either case we can apply our technique. That is, our "RE Tool" can be applied right throughout the system development life cycle.

We have described an RE framework as a small extension to an existing KE approach. RDR and FCA were used as subroutines within our RE system. For the example given, the only additional code needed for RE was about 100 extra lines of Visual Basic code added to a system that uses 2,500 - 3,000 lines of C code and 1,500 - 2,000 lines of Visual Basic code for the interface. We envisage that full implementation would require no more than a total of 500 extra lines. Thus, our RE extension to a KA tool will be a mere 10% extension (about 500 lines on top of 5000 existing lines).

FCA was used to build explanatory T-boxes from performance A-boxes. This approach is applicable to any representation which can be mapped into a decision table. However, during our discussion on resolution strategies (Figure 2), we noted that certain representations offered advantages. For example, when adding an attribute in a standard propositional rule base, the effects of this addition had to be checked all over the KBS. Such a check comes for free in RDR.

To instantiate our RE framework we have used the SISYPHUS III data. It has not been our goal to evaluate the accuracy of the KBS developed or their usefulness for classification or instruction, although we have shown that the MCRDR KBS are executable and argue that the concept lattices could act as a tutoring tool in helping the user learn about the domain. The goal of this paper has been to address an issue not stated explicitly in the problem statement for SISYPHUS III and that is how to manage multiple and conflicting sources of expertise. We have provided measures that show our resolution strategies can reduce the degree of conflict between multiple stakeholders. In tackling this problem, this approach has also addressed a drawback of standard RDR. RDR systems have been shown to be useful for single expert knowledge acquisition. In such a situation, RDR offers a good performance module, but a poor explanation module. However, in the case of multiple experts, an explanation system is required since experts must trade off their competing views. FCA allows us to build an explanatory T-Box from an A-box initialised by RDR.

References
Balzer, R. (1991) Tolerating Inconsistency Proceedings of 13th International Conference on Software Engineering (ICSE-13) Austin, Texas, USA, 13-17th May 1991, 158-165; IEEE Computer Society Press.

Brachman, R.J. (1979) On the Epistemological Status of Semantic Networks In Findler, N.V. (ed) Associative Networks: Representation and Use of Knowledge by Computers Academic Press-50.

Chandrasekaran, B. and Johnson, T. (1993) Generic Tasks and Task Structures In David, J.M., Krivine, J.-P. and Simmons, R., editors Second Generation Expert Systems pp: 232-272. Springer, Berlin.

Colomb, Robert.M. (1993) Decision Tables, Decision Trees and Cases: Propositional Knowledge-Based Systems Technical Report No. 266 Key Centre for Software Technology, Department of Compter Science, The University of Queensland, Australia.

Compton, P. and Jansen, R., (1990) A Philosophical Basis for Knowledge Acquisition. Knowledge Acquisition 2:241-257

Compton, P., Edwards, G., Srinivasan, A., Malor, P., Preston, P. Kang B. and L. Lazarus (1992) Ripple-down-rules: Turning Knowledge Acquisition into Knowledge Maintenance Artificial Intelligence in Medicine (4)47-59

Easterbrook, Steve (1991) Elicitation of Requirements from Multiple Perspectives PhD Thesis, Department of Computing, Imperial College of Science, Technology and Medicine, University of London, London SW7 2BZ.

Easterbrook, S. and Nuseibeh, B. (1996) Using Viewpoints for Inconsistency Management BCSEEE Software Engineering Journal January 1996:31-43.

Edwards, G., Compton, P., Malor, R, Srinivasan, A. and Lazarus, L. (1993) PEIRS: a Pathologist Maintained Expert System for the Interpretation of Chemical Pathology Reports Pathology 25: 27-34.

Finkelstein, A.C.W., Goedicke, M., Kramer, J. and Niskier, C. (1989) Viewpoint Oriented Software Development: Methods and Viewpoints in Requirements Engineering In Proceedings of the Second Meteor Workshop on Methods for Formal Specification Springer Verlag, LNCS.

Finkelstein, A., Gabbay, D., Hunter, A., Kramer, J. and Nuseibeh, B. (1994) Inconsistency Handling in Multi-Perspective Specifications IEEE Transactions on Software Engineering 20(8):569-578.

Gaines, B. R. and Shaw, M.L.G. (1989) Comparing the Conceptual Systems of Experts The 11th International Joint Conference on Artificial Intelligence :633-638.

Gil, Yolanda and Tallis, Marcelo (1997) A Script-Based Approach to Modifying Knowledge Bases In Proceedings of the Fourteenth National Conference on Artificial Intelligence and Ninth Innovative Application of Artificial Intelligence Conference AAAI Press/ MIT Press, Cambridge, Massachusetts.

Guha, T.V., and Lenat, D.B. (1990) CYC:A Mid-Term Report AI Magazine 11(3):32-59

Herlea, D.E. (1996) User's Involvement in the Requirements Engineering Process Proceedings 10th Knowledge Acquisition Workshop, Banff, Canada.

Kang, B., Compton, P. and Preston, P (1995) Multiple Classification Ripple Down Rules: Evaluation and Possibilities Proceedings 9th Banff Knowledge Acquisition for Knowledge Based Systems Workshop Banff. Feb 26 - March 3 1995, Vol 1: 17.1-17.20.

McDermott, J. (1988) Preliminary Steps Toward a Taxonomy of Problem-Solving Methods Automating Knowledge Acquisition for Expert Systems Marcus, S (ed.) Kluwer Academic Publishers, pp: 225-256.

Menzies", T. (1997) 35 Kinds of Knowledge Maintenance, Computer Science Technical Report. Submitted to the Knowledge Engineering Review. Available from URL http://www.cse.unsw.edu.au/~timm/pub/docs/97km.ps.gz.

Narayanaswamy, K. and Goldman, N. (1992) "Lazy Consistency": A Basis for Cooperative Software Development Proceedings of International Conference on Computer-Supported Cooperative Work (CSCW'92) Toronto, Ontario, Canada, 31 October- 4 November, 257-264; ACM SIGCHI & SIGOIS.

Nebel, B. (1991) Terminological Cycles: Semantics and Computational Properties In John Sowa (ed) Principles of Semantic Networks: Explorations in the Representation of Knowledge Morgan Kaufmann Publishers, Inc. California, 331-361.

Patil, R. S., Fikes, R. E., Patel-Schneider, P. F., McKay, D., Finin, T., Gruber, T. R. and Neches, R., (1992) The DARPA Knowledge Sharing Effort: Progress Report In C. Rich, B. Nebel and Swartout, W., Principles of Knowledge Representation and Reasoning: Proceedings of the Third Int. Conference Cambridge, MA, Morgan Kaufman.

Pirlein, T and Struder, R., (1994) KARO: An Integrated Environment for Reusing Ontologies European Knowledge Acquisition Workshop '94, Springer Verlag

Puerta, A. R, Egar, J.W., Tu, S.W. and Musen, M.A. (1992) A Mulitple Method Knowledge Acquisition Shell for Automatic Generation of Knowledge Acquisition Tools Knowledge Acquisition 4(2).

Ramesh, B. and Dhar, V. (1992) Supporting Systems Development by Capturing Deliberations During Requirments Engineering IEEE Transactions on Software Engineering 18(6):498-510.

Richards, D and Compton, P, (1997a) Finding Conceptual Models to Assist Validation Proceedings of AAAI'97 Verification and Validation Workshop Technical Report WS-97-01, July 28, Providence, Rhode Island. AAAI Press: Menlo Park, CA, pp:53-62.

Richards, D. and Compton, P. (1997b) Uncovering the Conceptual Models in Ripple Down Rules In Dickson Lukose, Harry Delugach, Marry Keeler, Leroy Searle, and John F. Sowa, (Eds) (1997), Proceedings of the Fifth International Conference on Conceptual Structures (ICCS'97), August 3 - 8, University of Washington, Seattle, USA, Lecture Notes in Artificial Intelligence, Springer-Verlag, Number 1257, Berlin pp:198-212

Rumbaugh, James (1994) Getting Started: Using Use Cases to Capture Requirements JOOP September 1994:8-23.

Schreiber, G., Weilinga, B. and Breuker (eds) (1993) KADS: A Principles Approach to Knowledge-Based System Development Knowledge-Based Systems London, England, Academic Press.

Schwanke, R.W. and Kaiser, G.E. (1988) Living with Inconsistency in Large Systems Proceedings of the International Workshop on Software Version and Configuration Control Grassau, Germany, 27-29 January 1988, 98-118;B.G. Teubner, Stuttgart.

Shadbolt, N., (1996)URL:http://www.psyc.nott.ac.uk/aigr/research/ka/SisIII

Shaw, M.L.G., (1980) On Becoming a Personal Scientist: Interactive Computer Elicitation of Personal Models of the World Academic Press, London.

Soloway, E, Bachant, J. and Jensen, K. (1987) Assessing the Maintainability of XCON-in-RIME: Coping with Problems of a very Large Rule Base Proceedings of the Sixth International Conference on Artificial Intelligence Vol 2:824-829, Seattle, WA: Morgan Kaufman.

Steels, L. (1993) The Componential Framework and Its Role in Reusability In David, J.M., Krivine, J.-P. and Simmons, R., editors Second Generation Expert Systems pp: 273-298. Springer, Berlin.

Strauss, A. (1978) Negotiation: Varieties, Contexts, Processes and Social Order Jossey-Bass Publishers, San Francisco, CA.

Thomas, K. (1976) Conflict and Conflict Management In Duneette (ed) Handbook of Industrial and Organisational Psychology Rand McNally College Publishing Co.

Wille, R. (1982) Restructuring Lattice Theory: An Approach Based on Hierarchies of Concepts In Ordered Sets (Ed. Rival) pp:445-470, Reidel, Dordrecht, Boston.

Wille, R. (1992) Concept Lattices and Conceptual Knowledge Systems Computers Math. Applic. (23) 6-9:493-515.