1. Introduction

It has become commonplace for critics of AI to note that relatively fewer expert systems are in widespread use than was predicted in the giddy atmosphere surrounding their initial development and early deployment in domains such as medicine. One reason typically given for the limited range of application of expert systems is their failure to gain acceptance by users. In particular, the inadequate approach to explanation found in most expert systems can be a significant impediment to user acceptance.

It stands to reason that a machine offering expert judgment in a given domain is more likely to find acceptance by those seeking its advice if it can explain its recommendations. Accordingly, an explanation capability should enable a user to get a complete, understandable answer to any sort of relevant question about the knowledge explicitly and implicitly embodied in a system's implementation formalism (e.g., rules, frames, or whatever). Unfortunately, the capacity of most current expert systems to explain their behavior (e.g., conclusions) is limited to either causal descriptions of the behavior of the performance environment's reasoning mechanism or to the display of canned text and graphics files. In this paper we will describe a new approach to explanation, which we will call participatory explanation.

2. Why is Explanation so Hard?

The typical article about explanation begins like this one; that is, it offers some compelling reasons for why an adequate explanation facility is crucial to the success of an expert system, and then laments the inadequacies of current explanation schemes (e.g., Jackson, 1986; Partridge, 1987). Much of the remainder of the paper is usually devoted to bemoaning the current state of the explanation art and speculating about what future AI (artificial intelligence) advances will be necessary before expert systems can offer explanations that take into account "the user's aims, objectives, hopes and fears, all with respect to the particular expertise of the AI system" (Partridge, 1987, p.4).

Any system able to meet the above goals, that is, to take into account the user's "aims, objectives, hopes and fears" will in essence be able to pass the Turing test for explanation -- thus, requiring a solution to the general AI problem before substantially improving the explanation capabilities of our systems. The sort of issues commonly raised in such discussions seem to be roughly equivalent to the general AI question (i.e., the whole ball of wax), and include such worthy foes as the problem of context, relevance, sufficiency (i.e., when to stop explaining) and other fundamentally difficult bugbears. Solve the AI problem, and you also will have solved the explanation problem, as well as many others. In the meantime, however, expert systems are coming into widespread use without adequate capacities for explaining their findings. This view of explanation places us in the somewhat peculiar position of waiting for the future advent of strong AI to make palatable the advice offered by the first broadly useful fruit (i.e., expert systems) of the AI research program.

3. Empowering Users to Construct Explanations

Our doubts concerning the near-term feasibility of meeting all the practical challenges facing traditional approaches to explanation do not imply that we assume that a fully intelligent explanation system embodying a passive mode of learning would resolve the problem. On the contrary, even if we were to postulate the existence of such a system, the epistemological objections raised here, as well as in the literatures of education and psychology, would remain unanswered. In short, we favor the design of explanation facilities that acknowledge and exploit the active role played by the learner/user in the process of meaning making. In response to many of the current and envisioned approaches to explanation, we are inclined to offer the slogan,

"too much instruction -- not enough human construction."

As we have noted elsewhere (Bradshaw, Ford, Adams-Webber, and Boose, 1993) designers of most expert systems leave little room for the active participation of users in their efforts to make meaning (i.e., construct an explanation) from the data presented to them in response to a request for explanation. The user is seen as a relatively passive receiver of the information intended to comprise an explanation.

In contrast, participatory explanation puts the user in an environment where he or she can assume an active role in the process of constructing his or her own explanation by freely exploring the domain model. A maxim that may help clarify our view of explanation is:

That which makes an utterance an explanation takes place in the ear (actually the mind) of the receiver of the explanation, not in the mouth of the provider of the explanation.

From this constructivist perspective, the user of any knowledge-based system is engaged in an ongoing cycle of observation, prediction, feedback and control. This constructivist outlook casts the designer of an explanation facility in the role of devising 'cognitive prosthetics' by means of which the system's users can stand on the shoulders of giants (i.e., the domain experts) in order to better understand the domain. Thus, a crucial question for knowledge engineers is to what extent can our explanation facilities support the users in their effort toward meaning making?

4. Foundations for Participatory Explanation

Elsewhere (Ford, Petry, Adams-Webber, & Chang, 1991; Ford & Adams-Webber, 1991; Ford, Bradshaw, Adams-Webber, & Agnew, 1992) we have elaborated a constructivist epistemology that can serve as a foundation for our efforts toward the development of user-centered participatory explanation systems. For the sake of brevity we will offer only a brief discussion here and direct interested readers to the above references. Although it is logically possible that there are an infinite number of different ways of interpreting some aspect of reality some conceptual models are more useful than others. Thus, in a sense, "experts" can be said to have developed significantly more useful models of the "reality" underlying a specific domain than has the ordinary practitioner. For example, domain experts can be viewed as having built-up repertories of working hypotheses, or "rules of thumb" (i.e., functional but fallible anticipations held with high confidence and uncertain validity) that guide their expert performance. We advocate attempting to make the expert's relatively superior model available to the community of users as a basis upon which the latter may directly improve their own performances, and also construct more useful explanations of relevant events (Pope & Gilbert, 1985). Concept maps (see Section 4.1) were developed for just this purpose.

4.1 Assimilation Theory and Concept Maps

Assimilation theory is a cognitive learning theory that has been widely applied to education (Ausubel, 1963; Ausubel, Novak, & Hanesian, 1978). Like Kelly's (1955) personal construct theory, it is based on a constructivist model of human cognitive processes. Specifically, it describes how concepts are acquired and organized within a learner's cognitive structure.

The concept map is assimilation theory's major methodological tool for ascertaining and representing what is known. In educational settings, concept mapping techniques have aided people of every age to examine many fields of knowledge. Much of the assimilation theoretic research to date has involved and exploited concept mapping (Novak & Gowin, 1984). In addition, concept maps are of increasing interest to those engaged in the process of knowledge acquisition for the construction of knowledge based systems (Ford, Stahl et al., 1991; Snyder, McNeese, Zaff, & Gomes, 1992). Essentially, concept maps provide context-dependent representations of a specific domain of knowledge within a set of concepts constructed so that the interrelationships among the included concepts are evident. In fact, concept maps have been shown to help students "learn how to learn" by making explicit their personally constructed knowledge and providing a structure for linking in new information. Concept maps offer a flexible framework for eliciting, representing, and communicating the emerging domain model. In this way, they are well suited to a participatory explanation paradigm in which the user explores the domain model built through the collaboration of the knowledge engineer, domain expert, and the modeling environment. Concept maps structure a set of concepts into a hierarchical framework. More general, inclusive concepts are found at the highest levels, with progressively more specific and less inclusive concepts arranged below them. In this way, concept maps display Ausubel's notion of subsumption, namely that new information is often relative to and subsumable under more inclusive concepts. All concepts at any given level in the hierarchy will tend to have a similar degree of generality. Relationships between concepts in a map represent propositions. Propositions form semantic units by linking together two or more concepts. Figure 1 shows a portion of a concept map produced by an expert in nuclear cardiology.

Figure 1. A portion of a concept map from the domain of nuclear cardiology.

In our approach to participatory explanation (see ICONKAT discussion in section 5.1), concept maps are an important mediating representation used to provide a hierarchically ordered, conceptual overview of the domain model arising from the collaborative efforts of the expert and knowledge engineer. The concept maps provide "knowledge landscapes" (essentially topographical maps) of the domain that comprise the organizational structure for entire domain model. It is into this semantic structure that other mediating representations (e.g., repertory grids, video, text, etc.) are linked during the knowledge acquisition phase. The ICONKAT approach to participatory explanation, described in Section 5.1, involves users constructing their own explanations while navigating their way through the linkages and nodes of the hierarchically organized concept maps comprising the domain model.

5. Knowledge Modeling

Participatory explanation, as do most approaches to providing advanced explanation facilities, requires access to an explicit model of the domain under consideration. This raises the issue of where the domain model comes from? Fortunately, it is possible to provide a modeling environment (e.g., ICONKAT) that supports the expert and knowledge engineer's efforts to collaborate in the construction of a domain model in such a way that the resulting model will be well-suited to participatory explanation. Instead of our arduously constructing a model of human expertise, and then throwing it away (upon translation into the syntax of the performance environment), an explanation facility also should exploit the model formed during the knowledge acquisition process. If not, then the implicit connections that establish the "logical" structure of the domain model may be lost, and as a result, much effort will be required to "put Humpty Dumpty back together again." This is essentially the task (i.e., reassembling Humpty) confronting those knowledge engineers who do not treat explanation as part of the knowledge acquisition process itself, and consequently, lack an adequate model in their attempts to construct explanation systems post hoc. Thus, one key to the design of explanation subsystems that are capable of more than shallow and/or mechanistic accounts is to recognize that the development of an explanation facility is a fundamental aspect of the knowledge acquisition process.

In Section 5.1, we discuss the specific explanation facilities of the ICONKAT (Ford, Stahl et al., 1991) knowledge acquisition tool. The ICONKAT approach to explanation reflects recent research in cognitive science (Yager & Ford, 1990), psychology (Kelly 1955; Pope & Gilbert, 1985) and education (Novak, 1977), which emphasizes an active and participatory approach to learning.

5.1 Participatory Explanation and ICONKAT

ICONKAT (Integrated Constructivist Knowledge Acquisition Tool) is a knowledge acquisition and representation system under evolutionary development at the University of West Florida. ICONKAT incorporates principles and techniques from both personal construct theory (Kelly, 1955) and assimilation theory (see Section 4.1). ICONKAT provides extensive interactive assistance to the domain expert and knowledge engineer in collaboratively modeling expertise. ICONKAT was used in the design and construction of NUCES: Nuclear Cardiology Expert System (Ford, Cañas, Coffey, Andrews, Schad, & Stahl, 1992). This is a large-scale expert system for the diagnosis of first pass cardiac functional images, a noninvasive radionuclide technique used to evaluate heart wall motion abnormalities.

In particular, ICONKAT supports the participatory explanation paradigm in which the domain model that emerges from the knowledge acquisition process is subsequently exported from the development environment to the delivery environment -- where it serves as the foundation of the explanation capability for the deployed system. ICONKAT's collaborative modeling environment exploits the expressiveness of concept maps to assist users in hierarchically organizing the various mediating representations (e.g., other concept maps, repertory grids, images, audio, video, documents) into browseable hypermedia domain models (see Figure 2). Interestingly, concept maps play a twin role in this process. First, concept maps are one of the principle means by which the expert and knowledge engineer represent knowledge about the domain. In particular, concept maps have proven excellent in eliciting and representing what the participants see as the knowledge landscape or topology at a given level of abstraction. Second, concept maps furnish a rich organizational framework that can serve as the interface to the domain model. Thus, while the expert and knowledge engineer collaborate in using concept maps to model the former's problem solving knowledge, they are also in essence building the structure of the interface that subsequent users will employ to explore the model when desiring an explanation. A session from NUCES (a medical expert system built in the ICONKAT environment) illustrates the participatory explanation approach (see Figure 2). When a user requests explanation, the performance environment is interrupted, and the user is switched into the context-sensitive explanation subsystem and conveyed to an appropriate location within the multi-dimensional space representing the model. From there, the user can assume an active role in the process of constructing his or her own explanation by freely exploring the conceptual model and browsing among a wealth of supporting objects (e.g., audio, video, documents, images, repertory grids, concept maps, rules, etc.). Users end their browsing as soon as they are confident that they have constructed an adequate explanation from the available information. Participatory explanation engages the user in an interactive process of observation, interpretation, prediction, and control.

(click on the image to see a large version)

Figure 2. A NUCES session illustrating the notion of participatory explanation.

The navigation problem, an important concern in participatory explanation, is largely ameliorated by use of concept maps as a guide to traversing the logical linkages among clusters of related objects (see the "Concept Map" window in Figure 2). Concept maps provide an elegant, easily understood interface to the domain model. A system of concept maps is interrelated by generalization and specialization relationships among concepts, which lead to a hierarchical organization. The explanation subsystem in NUCES provides a window that shows the hierarchical ordering of the various maps, highlights the current location of the user in the hierarchy, and permits movement to any other map by clicking on the desired map in the hierarchy (see the window "Concept Map Hierarchy" in Figure 2).

Depending on the location of the user in the domain model, he or she has different options to explore. At each node, the user can select from a menu of icons as shown in Figure 3. These correspond to text (a textual document), images, a popup menu of concept maps, repertory grids or video (implemented using QuickTime) related to the topic of the selected node. These icons will appear in various combinations depending on what information is available for a given concept. The "Concept Map" window in Figure 3 shows how the concepts (nodes) are populated with the icon menus illustrated in Figure 3. At any time, the user can backtrack by clicking on the "back-arrow" icon, as shown in the "Concept Map" window.

Figure 3. Close-up of the explanation icons.

This scheme provides the user great flexibility in navigating through related concepts, as well as guideposts in moving among the various sources of information available for a specific concept.

6. Summary

The main focus of this paper is to introduce participatory explanation -- a new approach to the design of explanation facilities for knowledge-based systems. This paper elaborates a constructivist theoretical foundation for participatory explanation. ICONKAT is presented as an example of a knowledge acquisition system that directly supports this approach to explanation. Further, a discussion of the explanation subsystem of NUCES, an expert system for nuclear cardiology, is presented as an example of a full-scale expert system that embodies the participatory approach to explanation.

7. References

Ausubel, D.P. (1963). The Psychology of Meaningful Verbal Learning. New York: Grune and Stratton.

Ausubel, D.P., Novak, J.D., & Hanesian, H. (1978). Educational Psychology: A Cognitive View (2nd Ed.). New York: Holt, Rinehart and Winston. Reprinted (1986). New York: Warbel and Peck.

Bradshaw, J.M., Ford, K.M., Adams-Webber, J.R., & Boose, J.H. (to appear Jan. 1993). Beyond the repertory grid: New approaches to constructivist knowledge acquisition tool development. In K.M. Ford & J.M. Bradshaw (Eds.), special issue on knowledge acquisition of International Journal of Intelligent Systems. Also to appear in K.M. Ford & J.M. Bradshaw (Eds.), Knowledge Acquisition as Modeling, New York: Wiley (1993).

Ford, K.M. & Adams-Webber, J.R. (1991). Knowledge acquisition and constructivist epistemology. In R.R. Hoffman (Ed.), The Psychology of Expertise: Cognitive Research and Empirical AI (pp. 121-136). New York: Springer-Verlag.

Ford, K.M., Cañas, A.J., Coffey, J., Andrews, E.J., Schad, N., & Stahl, H. (1992). Interpreting functional images with NUCES: Nuclear Cardiology Expert System. In M.B. Fishman (Ed.), Proceedings of Fifth Annual Florida AI Research Symposium (pp. 85-90). Ft. Lauderdale, FL: FLAIRS.

Ford, K.M., Petry, F., Adams-Webber, J.R., & Chang, P.J. (1991). An approach to knowledge acquisition based on the structure of personal construct systems. IEEE Transactions on Knowledge and Data Engineering, 3, 78-88.

Ford, K.M., Bradshaw, J.M., Adams-Webber, J.R., & Agnew, N.M. (to appear Jan. 1993). Knowledge acquisition as a constructive modeling activity. International Journal of Intelligent Systems.

Ford, K.M., Stahl, H., Adams-Webber, J.R., Cañas, A.J., Novak, J.D., & Jones, J.C. (1991). ICONKAT: An integrated constructivist knowledge acquisition tool. Knowledge Acquisition Journal, 3, 215-236.

Jackson, P.J. (1986). Introduction to Expert Systems. Reading, MA: Addison-Wesley.

Kelly, G.A. (1955). The Psychology of Personal Constructs, Vols. 1 & 2. New York: W.W. Norton.

Novak, J.D. (1977). A Theory of Education. Ithaca, NY: Cornell University Press.

Novak, J.D. & Gowin, D.B. (1984). Learning How to Learn. Ithaca, NY: Cornell University Press.

Partridge, D. (1987). The scope and limitations of first generation expert systems. Future Generation Computer Systems, 3, 1-10.

Pope, M. & Gilbert, J. (1985). Constructive science education. In F. Epting & A.W. Landfield (Eds.), Anticipating Personal Construct Psychology (pp. 111-127). Lincoln: University of Nebraska Press.

Snyder, D.E., McNeese, M.D., Zaff, B.S., & Gomes, M. (1992). Knowledge acquisition of tactical air-to-ground mission information using concept mapping. In Proceedings of the AAAI Cognitive Aspects of Knowledge Acquisition Session of the Spring Symposium (pp. 228-234). Stanford, CA: AAAI.

Yager, R.R. & Ford, K.M. (1990). A formal constructivist model of knowledge revision. In M.B. Fishman (Ed.), Proceedings of Third Florida Artificial Intelligence Research Symposium (pp. 154-158). Cocoa Beach, FL: FLAIRS.

Paper presented at FLAIRS 93: Sixth Florida Artificial Intelligence Research Symposium, Ft. Lauderadale, FL, April 18-21, 1993. Published in the Proceedings, pp. 111-115.