This position paper will state its premises, discuss requirements on the capabilities of interaction that these premises entail, and suggest two approaches that may aid in the satisfaction of these requirements. Finally, it will touch on the issue of the modeling of interaction raised by these considerations.
This paper will discuss the issues raised by the SEILLAC-II Organizing Committee from the viewpoint of the requirements on man-machine interaction which arise from the study, design, implementation and experience of knowledge based computer programs intended to function as consultants in particular fields of technical expertise. Quite impressive programs of this type have been constructed for such application areas as symbolic applied mathematics, analytical chemistry, and medical diagnosis and treatment. Work is currently in progress on the design and implementation of such programs for application to a variety of diverse fields. In order to function as a consultant, in contrast to a technical aid, these programs must share many of the properties we normally associate with the word consultant when applied to human practitioners. In particular, they must be generative, i.e., they must be capable of suggesting plans, designs, or options as well as evaluating those posed to them. They must function on the basis of knowledge of a particular field and be capable of employing this knowledge not only for the generation of options, but also for the explanation and defense of their suggestions. To make all this more difficult of attainment, they must often operate in situations in which either hard evaluational algorithms or sufficient data are lacking, so that their advice must be judgmental.
The knowledge contained in such systems is normally an entity quite aside from either the customary compiled program or input data. It is usually highly structured symbolic data which represents a model of the relationships between the data elements and the uses to be made of them, e.g., the procedures to be invoked automatically by changes in the data base. Some knowledge based systems possess symbolic models beyond that of the application domain. These may include models of the system's structure, models sufficient to support the understanding of some subset of a natural language, and models of context and dialog.
By habitability, we mean the naturalness and the ease of expression, as well as the lack of frustration and surprise during the interactive employment of such systems. As difficult as it is to pursue this habitability during the normal, consultative functioning of such systems, it becomes much more difficult during the knowledge acquisition (knowledge introduction and interactive knowledge debugging) phases. This is due to the lack of experience the practitioner may have with formal models - the application domain may not even be well codified - and, to be frank, the difficulty that even experienced artificial intelligence programmers have in predicting the changes in control structure introduced dynamically by changes in the knowledge base of complex programs.
We have chosen to discuss knowledge based consultant programs because that is the area in which we are working and in which we have been thinking most about the issues of interactive capabilities. Quite aside from personal concerns, it would seem propitious to discuss the demands posed by such system at SEILLAC-II for two reasons. First, knowledge based consultant programs are now emerging into practicality and it would seem reasonable to discuss the impact on issues of interaction that is likely to accompany this emergence. Second, most of the issues we shall raise are to be found among those raised by the SEILLAC-II Organizing Committee - but, in our case, pushed to an extreme of difficulty and complexity.
The position we are developing here will itself be an extreme one, arising for the most part out of what might be considered the unreasonable demands that we are placing on the habitability of knowledge based consultant programs. This is not meant to imply that compromise facilities are not valuable. We all must make such compromises and some such systems have achieved practical application. Nonetheless, we should recognize that the practitioners' demands for habitability are, while perhaps extremely difficult to satisfy, often very natural requests on his part. The degree to which they are satisfied or ignored may contribute as much to the fate of such a system as its competency. They should certainly have a place in the discussions at this conference.
We now pass to the requirements imposed by the above premises.
All computer programs remain suspect. A knowledge based consultant program, particularly one operating in a judgmental domain, will remain extremely suspect. It is therefore required that consultant programs be capable of explaining and defending their proposed plans or options. This capability for self-explanation based, as it must be, on the structure of the knowledge incorporated in the program, serves a second function beyond that of instilling and sustaining confidence in the program's competence. It is an absolutely essential component of the knowledge acquisition tool box, particularly with respect to knowledge debugging, i.e., the diagnosis and correction by a domain expert of judgmental errors on the part of the program. Experience with the MYCIN program, a consultant program dealing with the diagnosis and suggested treatment of bacterial infection, shows that the system need not only be capable of defending its recommendations, but also the reasonableness of the questions it poses to the physician.
Certainly the most difficult burden placed on the explanation capability by the demand for habitability is the requirement that the sense, level, and detail of an explanation should depend both on the context in which the question is posed and on a model the program maintains of the individual's sophistication, as a result of both an a. priori image of the user and the course of the current dialog.
The word command language is best applied to a user's interaction with a monitor. With certain consultant systems, the initiative is normally, in fact, with the program. The user must have the right to break this pattern, to answer a question with a question-and so must the machine! In a truly habitable system, the exchange of initiative would flow as naturally as the exchange of speakerhood in a conversation.
Another facet of this question of mixed initiative always arises almost immediately with the practical application of consultant systems, usually as a source of considerable user frustration. The user must have the ability-and few systems today provide it - to short circuit the normal functioning of the system. The user does not want to patiently respond while the system asks a long series of questions, each dependent perhaps on the previous responses. The user wants to make an initial dump of whatever facts he possesses and believes relevant. He wants, in effect, to be able to start the system up in an arbitrary initial state. He even wants to be able to retreat, to rethink and to change some of his previous responses, and to have the system behave gracefully while sustaining such a shock. The more expert: user will also want to short circuit the dialog through the introduction of advice and suggestions at any point in the transaction.
The most expert user must be capable of examining, understanding, and modifying the knowledge incorporated in the program. This implies, for one thing, the provision of a non-negligible portion of software-well tuned to both the particular application domain and the particular choice of knowledge representation - for the selection, display, and editing of the knowledge base, as well as configuration management tools for maintaining extended control of evolving versions of the program. Another type of tool, one of much greater inherent complexity, is also needed: a facility for predicting and explaining the consequences of proposed changes. In particular, the program should be able, if possible, to confirm the consistency of any change.
By now the deck is stacked. There would seem to be no way to satisfy the extreme requirements for habitability we have posed short of man-machine dialog conducted in a natural language. We have no intention to minimize, or even mediumize, the difficulties and disappointments that lie ahead in the task of creating computer programs capable of understanding natural language. Nor do we mean to imply that lesser ambitions are not worth pursuing nor that knowledge based systems with less sympathetic interfaces are not nor will they continue to be useful.
On the other hand, progress has been made and will continue to be made in research on machine understanding of natural language, quite aside from concern for the interactive demands of knowledge based consultant programs. Taking advantage of these would seem to represent the roost promising path to the fully habitable environment we have outlined above: the naturalness, the ease of expression, the acquisition of knowledge, the explanation of behavior at the appropriate level of detail, the graceful exchange of initiative - in short, the inference of user intent and choice of appropriate response.
In the long run, we suspect that some of the problems of natural language discourse may turn out to be, in fact, opportunities for man-machine interaction. Let us cite one example. The use of pronouns is generally regarded as an economy of language and their anaphoric disambiguation as a problem of natural language understanding. But recent linguistic study suggests that, when regarded at the level of a conversation, rather than that of an isolated sentence, the choice between the use of noun and pronoun bears considerable evidence as to the current focus of the dialog.
We would anticipate that graphical interaction will always supplement natural language for several specific purposes. The first is with respect to the presentation of the concrete objects in the domains of expertise of knowledge based consultant programs: mathematical equations and structures, chemical structures, maps, plans, designs, etc. A second use will be for the display of the knowledge incorporated in such systems. The fact that this knowledge is generally highly structured-the representation schemes bear such suggestive names as semantic nets, augmented transition nets, frame structures, not to mention the more familiar trees associated with hierarchically structured data or phrase structure grammars-implies that they will be much more legible in two dimensional presentations. Graphics will certainly play other auxiliary roles, e.g., the presentation of histograms representing the statistical consequences of experimental changes in the knowledge base.
In a sense, we have abrogated this question. By deriving such complex requirements from the goal of habitability in knowledge based consultant programs, we have essentially sublimated the question of models of interaction to that of models of discourse in natural language. The problems become those of understanding context, focus, exchange of initiative, interruption and reprise, and finally, of intent. Perhaps the linguists will help us.
