Proceedings of the 33rd Hawaii International Conference on System Sciences - 2000
A Situated Cognition Approach to Conceptual Modelling Debbie Richards Department of Computing School of Information and Communication Sciences Macquarie University, Sydney, Australia, 2109
[email protected] Abstract Conceptual modelling is an important task in the development of computer systems regardless of whether they are data based or knowledge based systems (KBS). Most approaches see conceptual modelling as a prerequisite to the capture of data or knowledge. These approaches implicitly assume that it is possible to capture and validate a “good” model. However, modelling is difficult, time-consuming and error-prone. The approach described in this paper is based on a situated view of cognition and the premise that it is not easy to capture or evaluate a conceptual model. The alternative offered is based on the use of cases, Ripple-Down Rules (RDR) and Formal Concept Analysis (FCA). Cases are used to motivate the capture of rules in a simple user-driven manner. RDR is used as the knowledge acquisition and representation technique. The propositional rules are then interpreted as a binary formal context and a complete lattice is automatically generated using FCA. In this way, contrary to mainstream approaches, we begin with an assertional KBS and later derive a terminological KBS. Cases ground the KBS in the real world and provide the context in which the knowledge applies. The ease with which the knowledge is acquired and maintained allows for the continual evolution of the KBS in keeping with the notion that knowledge is continually evolving and “madeup” to fit the situation.
1. Introduction Conceptual modelling is an important task in the development of computer systems regardless of whether they are data or knowledge based systems (KBS). In each of these systems, the designer attempts to capture the main concepts of the domain relevant to that system. The goal of the model is to assist in the acquisition and management of knowledge and/or data for the domain. Much research in these areas makes use of some form of an abstraction or type hierarchy. In object oriented systems this hierarchy is obvious and a key to the success of the technology. In KBS research this hierarchy is often
termed an ontology. While definitions of the word “ontology” abound, on a very general level an ontology can be seen as an explicit specification of, “some abstract, simplified view of the world”[31, p.3]. The research reported in this paper is particularly concerned with KBS but is seen to be applicable to many types of computer systems in so far as the focus is on conceptual modelling. Modelling, however, is a time-intensive, complex and error-prone task. Approaches which rely on the development of a good model as a prerequisite to the capture of knowledge or data implicitly assume that it is in fact possible to acquire and validate “good” models. An alternative approach is described in this paper which is based on a situated view of cognition and the premise that it is not easy to capture or evaluate a conceptual model. The approach offered combines the use of cases, rippledown rules (RDR) [4] and formal concept analysis (FCA) [35]. Cases are used to motivate the capture of rules in a simple user-driven manner. RDR is used as the knowledge acquisition (KA) and representation technique. The propositional rules are then interpreted as a binary formal context and a complete lattice is automatically generated using FCA. In this way, contrary to mainstream approaches, we begin with an assertional KBS and later derive a terminological KBS. Cases ground the KBS in the real world and provide the context in which the knowledge applies. The ease with which the knowledge is acquired and maintained allows for the continual evolution of the KBS in keeping with the notion that knowledge is continually evolving and “made-up” to fit the situation. We look first at the limitations of conceptual models and implications of situated cognition on the approach taken. In Section Two, RDR and FCA are introduced. In Section Three we look at a case study which describes how RDR and FCA are used to develop and compare conceptual models. Section Four considers other and further research in this area. Finally, we conclude with consideration of the role of retrospective conceptual modelling.
0-7695-0493-0/00 $10.00 (c) 2000 IEEE
1
Proceedings of the 33rd Hawaii International Conference on System Sciences - 2000
1.1 The Problem with Conceptual Modelling
2.1 An Introduction to Ripple Down Rules
The term “conceptual modelling” suggests that we are dealing with something that is abstract and likely to be hard to extract, understand and represent and that the end result will be an approximation at best. A model can be seen as “a description and generator of behaviour patterns over time, not a mechanism equivalent to human capacity” [3, p. 89]. Further, models have been shown to vary between experts in the same domain and even over time for the same individual [10]. If we take a situated view of cognition, human thought and action are inextricably connected and affected by the context. It is not just the external environment that will affect the context but that thinking itself modifies further action and context occurs at a conceptual level that exists within a social setting [3]. The reliability of models is further questioned if we adopt Popper’s view [23] which implies the impossibility of deduction in an ultimate sense. It is no wonder that conceptual modelling is difficult and error-prone. Furthermore, the emphasis of most KBS approaches on building terminological KBS as a prerequisite for the capturing of expertise is not consistent with the way that experts behave: “If we take Heidegger and Maturana seriously, we see that experts do not need to have formalised representations in order to act” [36, p. 99] and “reflection and abstraction are important phenomena, but are not the basis for everyday action” [36, p. 97]. Winograd and Flores argue that it is the ability to act spontaneously that makes an expert an expert - “The essense of our intelligence is our throwness not our reflection” [36, p.99]. Throwness [16] refers to the common human experience of acting in response to being thrown into a situation rather than reflecting first and then acting. From this body of work it appears that most expert action is reflexive. Since conceptual models are unreliable and difficult to capture an approach based on behaviour may be preferable to one that requires the expert to articulate their thought processes. Knowledge acquisition, maintenance and inferencing using RDR are simple activities which are provided in reflexive modes that do not require introspection. Using FCA the higher level model is uncovered later.
RDR were developed to address many of the issues raised by situated cognition [26]. One issue is the need to support ongoing maintenance which is based on the view that knowledge is never complete. Another issue is the need to provide the context in which the knowledge applies. This is done through the use and storage of cases. Another issue is the need to give the user control and provide direct interaction with the system. An awareness of the imperfections of models and acceptance that we as yet do not understand how an expert thinks has resulted in an approach that focuses more on the behaviour of the expert as they interact with a case. It was observed that experts are competent at assigning a conclusion to a case or scenario. When asked to explain that conclusion they tended to offer a justification which differed according to the case and the audience [5]. Single-classification RDR were developed first and can be defined as a triple [29] where X are the exception rules and N are the if-not rules. When a rule is satisfied the exception rules are evaluated and none of the lower rules are tested. The major success for this approach has been the Pathology Expert Interpretative Reporting System (PEIRS) [8], a large medical expert system for pathology laboratory report interpretation built by domain experts with minimal intervention of a knowledge engineer. PEIRS is one of the few medical ES that have actually gone into routine use. It went into operation with 198 rules and expanded over four years to over 2000 rules, covering 12 different pathology tests. A total of approximately 100 hours were spent on KA. The development of 10 rules per hour for RDR is outstanding compared to industry standards of only 2 or 3 rules per day [7]. Multiple classification RDR (MCRDR) have more recently been developed to handle classification tasks where multiple independent classifications are required [18,19]. This method builds n-ary trees and consists only of exception branches. A better description may be sets of decision lists joined by exceptions. In contrast to single classification RDR all rules attached to true parents are evaluated against the data. Figure 1 shows an example MCRDR built using Cendrowska’s [2] lens data with two levels of decision lists. The label “Corners” in the diagram refers to the cases associated with that rule and will be discussed further on. An MCRDR is defined as the quadruple , where P is the parent rule, C are the children/exception rules and S are the sibling rules within the same level of decision list. Every rule in the first list is evaluated. If a rule evaluates to false then no further lists attached to that rule are examined. If a rule evaluates to true all rules in the next list are tested. The list of every true rule is processed in this way. The last true rule on each path constitutes the conclusions given.
2. Introducing the theories The main focus of this paper is on conceptual modelling and how RDR and FCA can be used for this purpose. While both theories are widely published a description of both will be given here as they have provided the foundation for the work. Due to space limitations, it is only possible to provide a brief introduction to the theories. The interested reader is invited to explore some of the references given for a more detailed description.
0-7695-0493-0/00 $10.00 (c) 2000 IEEE
2
Proceedings of the 33rd Hawaii International Conference on System Sciences - 2000
Rule 1: If tear_production=normal and astigmatic=no Then Lens=soft Corners (1,2) Rule 0: If 1=1 Then Lens=none
Rule 4: If age=presbyopic and prescription=myope Then Lens=none Corners (5)
Rule 2: If astigmatic=yes and tear_production=normal and presciption=myope Then Lens=hard Corners( 1,2,4) Rule 3: If astigmatic=yes and tear_production=normal Rule age=young 7: and If d,g Lens=hard then Cls 5 Then Corners(1,3)
Figure 1. An MCRDR KBS for the Contact Lens Prescription Domain [2]. KA using RDR involves the expert being shown a case with a system assigned conclusion. If the expert agrees with the conclusion they continue to the next case. If they do not agree they assign a different conclusion and pick some features of the case which form the rule conditions. The features are attribute value pairs. The case that prompts a new rule to be added becomes stored in association with the new rule and is known as the cornerstone case. When a new rule is added, the cornerstone case of the rule that gave the misclassification is shown to the user and they must pick some features which distinguish the two cases. Rules are never deleted or changed. Each new rule is a modification of a previous rule. This may appear inefficient, but studies have shown the exception structure representation to be at least as compact as KBS built using the machine learning algorithms C4.5 and Induct [5, 6, 21]. In single classification RDR only one case is associated with each rule. In MCRDR there may be multiple cases that must be distinguished from the current case. In the KA approach developed [18], the expert is presented with one cornerstone case at a time. The expert constructs a rule to distinguish the new case from the first case presented and then each case in the cornerstone list is evaluated to see if it is also distinguished by the new rule. If a case is satisfied by the rule the expert must add extra conditions to the rule to distinguish this case. This continues until all related cornerstone cases are distinguished. Remarkably the expert provides a sufficiently precise rule after two or three cases have been seen [19]. The decision of where to add a new rule to the KB is affected by the design of the user interface, user preferences and the situation. If rules tend to be added to the top level the domain will be covered more rapidly but there may be greater errors. If rules are added to the end of pathways less cases will be seen so there will be less errors but slower domain coverage [18]. New cases may be misclassified in one of three ways: one or more of the
conclusions are incorrect, one or more conclusions are missing or a combination of incorrect and missing conclusions. The user may decide to stop an incorrect conclusion instead of replacing it with a new conclusion. This is achieved by adding a rule which has a null conclusion in the same way as adding other types of rules. The ease with which KA and maintenance (these are one and the same in RDR) are performed by the domain expert is a major strength of RDR and the reason why it was chosen as the method of acquiring the initial model. However, the RDR model only provided a low-level view of the knowledge. This led to the addition of techniques from FCA.
2.2 An Introduction to Formal Concept Analysis Formal concept analysis draws on ideas from lattice and order theory [35]. A concept in FCA is comprised of a set of objects and the set of attributes associated with those objects. Alternatively, we can view the objects and attributes as entities and properties, respectively [32]. Knowledge is seen as applying in a context and can be represented as a crosstable and defined as a formal context. A formal context is a triple (G,M,I) where G (for Gegenstande in German) is the set of objects which forms the extension of the concept, M (for Merkmale in German) is the set of attributes which forms the intension of the concept and I is a binary relation connecting G and M. In the crosstable shown below, the rows are objects and the columns are attributes. An X indicates that a particular object has the corresponding attribute. Figure 3 shows a crosstable. Using the notion of a galois connection, formal concepts are found by determining the set of attributes shared by a set of objects or conversely the set of objects which share a set of attributes. More formally, the derivation operators: X⊆G: X Y ⊆ M: Y
a X′ :={m∈ M | gIm for all g ∈ X } (F.1) a Y′ :={g ∈ G | gIm for all m ∈ Y} (F.2)
0-7695-0493-0/00 $10.00 (c) 2000 IEEE
3
Proceedings of the 33rd Hawaii International Conference on System Sciences - 2000
are used to construct all formal concepts of a formal context, by finding the pairs (X′′,X′) and (Y′,Y′′). We can obtain all extents X′ by determining all row-intents {g}′ with g ∈ G and then finding all their intersections. Alternatively Y′ can be obtained by determining all column-extents {m}′ with m ∈ M and then finding all their intersection. This is specified as:
X′ =
I {g} ′
(Formula 3)
I {m} ′
(Formula 4)
g ∈X
Y′ =
m∈Y
Having found the set of formal concepts we can order these using the subsumption relation ≥ on the set of all concepts formed such that (X1,Y1) ≤ (X2,Y2) iff X1 ⊆ X2. By finding the predecessors and successors of each concept we can produce a visualisation of the concepts as a complete lattice, as shown in Figure 4. For a family (Xi,Yi) of formal concepts of K the greatest subconcept, the join, and the smallest superconcept, the meet, are respectively given by:
∨ (X ,B ):= ((U A )" , I B ) ∧ (X ,B ):= (I A , (U B )" ) i ∈I
i
i
i ∈I
i
i ∈I
i
i
i
i
i ∈I
(Formula 5)
i ∈I
i
(Formula 6)
i ∈I
RDR and FCA share a number of views including the beliefs that knowledge applies in a context and that KA is a task that is best performed directly by experts. In both approaches KA is reduced to the task of classifying objects (cases) and the identification of the salient features. In FCA, KA begins with the elicitation of a crosstable from which the concepts derived can be used to generate implications. The implications generated are shown to the user who is asked to say whether they agree or disagree with the implication. If the user does not agree they are asked to offer a counterexample. This study starts from the opposite direction by using the rules in the MCRDR KBS as the input into a formal context. The reason for this is twofold. Firstly the purpose of using FCA was to uncover higher models in rules that had already been acquired using MCRDR. Secondly, it was felt that the RDR approach to KA was probably less demanding for experts than the development of crosstables, the analysis of the generated implications and the offering of counterexamples which is required by the FCA approach to KA. The combined approach allows the strengths of both approaches to be combined. That is we get the benefit of easy KA and maintenance in an executable system from RDR with the more structured and deeper model provided by FCA. The combined approach is now described in a case study.
3. Retrospective Conceptual Modelling In this section, retrospective conceptual modelling is demonstrated. Minimal analysis of the domain is performed. Essentially, we produce a model of primitives by developing an RDR KBS from the cases. The primitive model is used by FCA to generate an abstraction hierarchy. The hierarchies can be compared using Gaines and Shaw’s [10] four Quadrant model. The data used in this exercise are taken from a real study that was performed to develop a combined KBS which represented the current state of what was known about a new type of grass.
3.1 The domain The domain chosen concerns the adaptation and management of the Lotus Uliginosis cv Grasslands Maku for pastures in the state of New South Wales [17]. This domain is particularly appropriate for examining the development and comparison of conceptual models since little was known regarding this crop. The goal of building the combined KBS was to gather knowledge about the conditions (such as weather and soil) that were favourable for this crop, to determine how well the crop performed as compared with other alternative crops and how best to manage this crop.
3.2 Acquiring the Knowledge An RDR system and an empty knowledge base were given to four independent agricultural advisers. These advisers represented local groups of farmers and agribusiness people who were involved with “co-learning” about the new crop which will be referred to as Lotus. RDR had been chosen as they were seen to be a suitable method of capturing emerging knowledge due to their support of rapid and easy KA and maintenance performed by domain experts. The four KBS developed will be referred to as Lotus 1, 2, 3 and 4. Figure 2 shows the MCRDR KBS developed for Lotus 1.
3.3 Generating the Hierarchy The first step was to use the rules to generate a formal context. The RDR KBS was converted to a flat form by traversing the list structure for each rule picking up the conditions from the parent rule until the top node with the default rule was reached. From this flattened KBS the user can choose either the whole KB or a more narrow focus of attention from which to derive a formal context. When the whole KB is chosen the rules and rule conditions form the extents and intents, respectively. Such a global view is only feasible for small, if not very small, KBS. This is due to the exponential nature of the algorithm and the problems associated with displaying and handling large graphs. More efficient algorithms can be found in [13].
0-7695-0493-0/00 $10.00 (c) 2000 IEEE
4
Proceedings of the 33rd Hawaii International Conference on System Sciences - 2000
1 0 11 0 %NC000 : 1 = 1 2 1 8 0 %MAKUU : (LOW_PH = YES) 3 1 0 2 %RYEGR : (RYEGRASS >= 15) 4 2 0 0 %NC000 : (RYEGRASS >= 15) 5 1 0 3 %SALIN : (SALINE = YES) & (WATERLOG = YES) 6 1 0 5 %SALIN : (LOW_PH = NO)& (WATERLOG = YES) 7 1 0 6 %SHARN : (LOW_PH = YES) 8 1 0 7 %MAKUU : (LOW_PH=NO) & (LOW_FERTILITY = NO) 9 1 0 9 %CLOVR:(WATERLOG=NO) & (LOW_FERTILITY =YES) 10 1 0 10 %MAKUU : (LOW_PH = YES)
Figure 2: The MCRDR KBS for Lotus 1. The numbers on each line are the rule number, the rule number of the parent, the rule number of the first child and the rule number of the next sibling, respectively. The five digit conclusion code is shown next and starts with a % for identification purposes. After the colon is a list of conditions joined by the conjunction operator &. This is only one of a number of files that make up the KBS. There are other files that store information such as links to cases, comments, user-defined functions and conclusion code definitions.
To address the problem of dealing with large contexts, there are currently 13 different ways a context may be derived. The two main methods are choosing a conclusion or a rule. If a conclusion is chosen, all rules using that conclusion are selected and added as objects to the set G, forming the extents of the context. As each extent is added the conditions of the rules are added to the set M of attributes to form the intents of the context, first checking to see if any attributes have already been added by previous rules. Where the relation I holds, that is object g has attribute m, a cross is marked in the appropriate row and column. If the user chooses a particular rule then that rule is added as the first object with the rule conditions as the initial intension. Every condition in each rule in the flattened RDR rule base is searched for a match on the Rule/Conclusio n 1-%NC000 2-%MAKUU 3-%RYEGR 4-%NC000 5-%SALIN 6-%SALIN 7-%SHARN 8-%MAKUU 9-%CLOVR 10-%MAKUU
1=1* X X X X X X X X X X
LPH =Y
RYE >=15
SAL =Y
initial set of attributes. If a match is found, that rule is added to the extension and all new attributes (conditions) found in the matching rule are also added to the intension. Figure 3 shows the crosstable for the KBS in Figure 2. Since each Lotus KBS had from 11-18 rules it was not necessary to shorten [12] the context. Treating the rule conditions as boolean attributes, is similar to the technique known as conceptual scaling [14] which has been used to interpret a many-valued context into a (binary) formal context. A many-valued context, such as that represented in an MCRDR KBS, is a quadruple (G,M,W,I) where I is a ternary relation between the set of objects G, the set of attributes M and the set of attribute values W (merkmalsWerte in German). Essentially, each attribute is treated as a separate formal context with the values as attributes associated with each of the original objects. A scale is chosen, such as a nominal scale (=) or an ordinal scale (≥), to order these attributes. From the many contexts, one for each attribute, the concepts are derived. The crosstable shown in Figure 3 was then used to construct all formal concepts of the formal context, using the process described in Section 2.2. To allow drawing of the Hasse diagram it was necessary to compute the predecessors and successors of each concept. Predecessors were found by finding the largest subconcept, the join (Formula 5), of the intents for each concept. Successors were found by finding the smallest superconcept, the meet (Formula 6), of the intents. The successor list was used to identify concepts higher in the diagram, the parents, and the predecessor list identified concepts lower in the diagram, the children. As Wille [34] points out, there is not one fixed way of drawing line diagrams and often a number of different layouts should be used because concepts can be viewed and examined in different ways depending on their purpose and meaning. The user may also move a node anywhere they like providing the node is not moved higher than any of its parents or lower than any of its children. WA T=Y
LPH =N
LFE =N
WAT =N
LFE =Y
X
X
X X
X X X
X X
X
X X
X
X Figure 3: A Crosstable for the Lotus 1 KBS.
LPH=LOW_PH, RYE=RYEGRASS, SAL=SALINE, WAT=WATERLOG, LFE=LOW_FERTILITY. * 1=1 is the default rule condition which always evaluates to true. Figure 4 provides a description of each of the conclusion codes.
0-7695-0493-0/00 $10.00 (c) 2000 IEEE
5
Proceedings of the 33rd Hawaii International Conference on System Sciences - 2000
%CLOVR- Maku Ok, you could consider clover plus fertiliser %DEEP1-Lotus sown too deep - establishment failure likely.Maku should be seeded no deeper than 15mm or it will have difficulty in emergence. %DRYSO-Not enough moisture at sowing. %FLOOD-Maku is killed by 3 or more days of flooding if water is still and high temp %KENYA-In high rainfall and on the better soils Kenya clover may provide a higher quality pasture. %LIGHT-Maku unsuited to light quick drying soil %MAKUU -Better suited to Maku than to Clover %MAKU2-Maku should persist %MAKU3-Maku Ok, but flooding with limit persistence %MAKU4-Maku Ok only at higher sowing rates %MAKU5-Maku at rainfalls above 1250mm %MANAG-Manage grazing to control grass competition in spring.%RENOV-Pasture renovation with oats can be too competitive for lotus. %RYEGR-Ryegrass density is too high for lotus to compete %SALIN -Maku or try strawberry clover Maku OK. %SCARA-Persistence of Maku may be limited by scarabs %SEASN-it is safer to sow lotus in autumn. %SHALW-Maku is unlikely to persist through a prolonged dry period.%SHARN-Maku or try Sharnae/Goldie %SUMGR-To avoid excessive competion from summer grasses, manage grazing to ensure grasses are kept down during the summer period.%SUMM1-grass competition is excessive for Lotus - keep grass down in summer
Figure 4: Descriptions for the Lotus Conclusion Codes
3.4 Comparing the Conceptual Models There are a number of ways the concept lattice for each KBS can be compared. One approach is to visually compare each lattice to determine what differences and
similarities exist. While it is noted that comparison using FCA is quite simple it was desirable to automate the comparison process and to offer some assistance to the user by focusing the comparison. To this end, MCRDR/FCA will automatically generate a lattice from the combined KBS choosing to include only shared concepts OR only different concepts. Such an approach immediately shows what common ground exists and what differences need to be discussed. To further assist in comparison by focusing the comparison, a third approach is shown in Figure 5 which is the result of selecting certain aspects of the combined KBS. In Figure 5 all attributes concerned with water were used as the basis of generation of the formal context. These attributes included the key words: Moisture, Water, Flood and Rain. A total of ten concepts were generated. Concept labelling has been reduced to minimise screen clutter. Objects and attributes belonging to a concept are reached by descending and ascending paths, respectively. Thus Concept Number 8 includes the attributes {LOW_PH=NO, WATERLOG=YES} which are shared by the objects which rule 6 in each KBS represents. These objects conclude "%SALIN- Maku or try strawberry clover". From the lattice we can see that all viewpoints agree on a number of concepts. The KBS where each object (rule) has originated from is shown as L1-L4. First of all, they all agree (see concept 9) that the conclusion should be %SALIN when WATERLOG=YES and SALINE=YES or when WATERLOG=YES and LOW_PH=NO. They also all agree that when WATERLOG=NO and LOW_FERTILITY=YES the conclusion should be "%CLOVR – Maku Ok, you could consider clover plus fertiliser". However, Lotus 2 believes that if RAINFALL=YES then the conclusion should be
Figure 5: The Concept Lattice derived from all rules in all four Lotus KBS which use the key words: rain, water, moisture or flood.
0-7695-0493-0/00 $10.00 (c) 2000 IEEE
6
Proceedings of the 33rd Hawaii International Conference on System Sciences - 2000
"%MAKU2- Maku should persist" (concept 4) and if FLOODING>=3 then %FLOOD-Maku is killed by 3 or more days of flooding if water is still and high temp (concept 5) is concluded. Lotus 3 recommends "%MAKU3- Maku Ok, but flooding with limited persistence" if SUMMER_FLOOD>=3. Lotus 4 recommends "%MAKU5- Maku at rainfalls above 1250mm" if MOISTURE