User Modelling

USER MODELLING PANEL

D. Sleeman (Stanford: Moderator), Doug Appelt(SRI), Kurt Konolige(SRI), Elaine Rich(MCC), NS Sridharan(BBN Labs) & Bill Swartout(ISI).

INTRODUCTION

ln various sub-areas of Al we talk about "tailoring" the system's response to the user. NL systems and Tutoring systems being two prime examples. Additionally, some discussion of this issue arises in building explanation facilities torExpert Systems.

• e How explicit are the user models even in systems which are able to adapt to the user?
• How do they achieve this tailoring? How similar are the techniques used?
• How do such user models diller trom the plans interred in planning systems?
• How deep/knowledgeable do \Jser models need to be?
• How is this sophistication dependent on the type of interaction (superficial conversation versus diagnostic/ tutorial), the goal of the dialogue, the nature of the domain etc?

ln this panel we will review many of the areas in which some form of user model is used, Iook at commonalities of approaches, and seek to characterize when a particular approach is appropriate.

USER MODELLING: SOME APPROACHES Elaine Rich

User modelling straddles the boundary between artificial intelligence and data base technology. lt has all of the problems that each of these areas possesses; we hope it will also be able to draw on both areas for solutions. This double dependency arises from the interaction between the two main subproblems that user modelfing must address:

• How can models of users (their knowledge, goals, etc.) be inferred from their behavior and used in reasoning to improve the performance of a target system? This is where A.l. comes in.
• How can models of a large number of users be maintained efficiently so that each is available when necessary but system pertorrnance does not degrade

even if the user population is _very large? This is where data bases are irnportant.

User rnodels capture rnany kinds of information about users. Two irnportant dirnensions that characterize this intorrnation are shown in the following chart:

Square 2 does not make much sense, but each ot the other squares poses specific problems that user-modelling systems must address.

When usermodelfing is looked at from the A.l. point of view, the following issues emerge:

• How can specitic user plans be interred trom behavior? (This relates to square 1.) Doing this requires a system that is itself capable ot forming plans but it also, since it is a diagnosis task and not a design task the way most planning problems are, requires a sophisticated matehing procedure so that the plan that the user has selected can be isolated from other possible plans.
• How can user knowledge and planning strategies be inferred from behavior? (This relates to square 3.) People's knowledge and their problem-solving strategies change quite slowly over time and should be remernbered trom one session to the next because they may substantially influence both the way the user will behave and the way the system should behave tor maximum effectiveness.
• How can general knowledge and planning procedures be represented and used effectively? (This relates to square 4.) This is the standard A.l. question.

When user modelfing is tooked at trom the database point of view, the following issues emerge:

• How can information about a large number of users be stored most efficiently? (This relates to square 3.) ls it more efficient to store each model separately or to store models as differences from some canonical model? ls it better to organize the model areund a particular user, clustering Iogether everything known about that person, or is it better to arrange the model areund a particular topic or piece of knowledge, dustering Iogether what is known about all users with respect to that issue?
• How can a Iot of detailed knowledge about a particUiar session be collapsed into a concise description of the knowledge that may be useful for later sessions? (This relates to square 3.)

STUDENT MODELS in INTELLIGENT TUTORING SYSTEMS D. Sleeman

The field of Intelligent Tutering Systems identified the need for having a model · a database . which summarized the student's actions some time ago. The aarliest adaptive CAI programs often represented the student's Ievei of sophistication by a scalar value. SCHOLAR, a program which discussed the geography of South America, was the first to use a more sophisticated representation . namely a semantic network. Essentially, the knowledge of the domain was represented as such a network and each node had a numerical value associated with it indicating the likelihood that a particular student knew the knowledge associated with the node. This type of model is referred to as an overlay model. A differential model which simply reports the differences between an expert's and the student's knowledge was introduced in the WEST system.

All these systems assumed that the student's knowledge was merely a subsei of the expert's. Recent studies in Cognitive Science have shown this is frequently not a valid assumption, and so models which allow both the correct and incorrect knowledge to intermingle have been introduced. Perturbation models have been used by Brown & Burton in their DEBUGGY system to modal student's errors with Arithmetic tasks, and by Sleeman in LMS/PIXIE to capture student's knowledge of Algebra.

An important aspect of these latter models is that they are process models · and so can be executed by an appropriate interpreter · thus enabling them to be used predictively. 8oth DEBUGGY and PIXIE address the issue of interring models by observing the student's performance on a series of tasks. Technical issues addressed by these systems include how to make the search computationally tractable, and how to overcome noise (i.e., spurious responses). Additionally, PIXIE is addressing the issue of how to remove the closed-world assumption · making the systems truely responsive to the student's input. Currently, modelling systems merely search · an albeil very large · model space generated by combining more primitive components (in PIXIE's case of correct and incorrect rules).

EXPLANATION & the ROLE of the USER MODEL: HOW MUCH WILL IT HELP?

Bill Swartout

There seems to be a growing consensus among researchers in explanation and text generation that a solid, detailed user model (if we only knew how to build it) would significantly improve the kinds of explanations and texts we can produce mechanically. Currently proposed system designs often call for a detailed user model that expresses what facts the system believes the user knows, how he likes to have information presented, and so forth. ln such designs, presentation strateg'ies use the model to select just the right thing to present to a user. ls such a detailed model feasible? Do people seem to have detailed knowledge of their listeners? This approach may place too much emphasis on the user model. lt often seems that people do not have detailed knowledg.e of their Iistenars but instead rely on general, stereotyp1cal knowledge and an ability to alter their explanation tactics when the listener appears not to understand. I would like to suggest that an explanation system that allows for feedback from the user about the Understandability of explanations and that relies on a general user model expressing knowledge of stereotypes might be more feasible that one that depends on a detailed user model.

THE ROLE OF USER MODELLING IN LANGUAGE GENERATION & COMMUNICATION PLANNING

Doug Appell

The analyses of Searle and Grice clearly demonstrate that communication is a process of intended recognition of intention, whereby the speaker formulates utterances with the intention that the hearer use that utterance to understand the speaker's intentions that the hearer hold some different propositional attitudes as a result of understanding the utterance. This intention recognition property is essential to communication . .. if it is absent, then whatever activity is going on is something other than communication.

lf a user perceives natural language being used as input and outpul to a system, it is very natural for him to assume that it is being used as a medium ol communication, much the same as people use it among themselves. Therefore, there is a very streng tendency for the user to impute intention recognition capabilities to the system and to assume that it is taking his own intentions into account. Of course, most users of currently available natural language interfaces soon learn that this is not the case. The objective of research in communication planning is not so much being able to construct ever more complex sentences involving increasingly difficult semantic concepts, but rather to understand the processes of intention communication and recognition weil enough to enable a system to participate in a natural dialogue with its user.

Therefore a system that plans communication must have a very detailed rnodel of the user. There are a large number of alternative means of representing the beliefs and intentions of agents, and the requirements of communication planning do not dictate what form such a representation must take, but rather dictates a set of requirements about what kinds of reasoning must be done. The following is at least a partial Iist of the representation and reasoning capabilities necessary for communication:

• The ability to represent Believe(A, P), Believe(A, -P), -Believe(A, P).
• The ability to represent all of the above with respect to mutual belief.

• The ability to represent all of the above with respect to intention.
• The ability to deduce for any P whether or not A believes P, and similarly for mutual belief and intention.
• Given an individual, reason about what is believed or mutually believed about it.
• The ability to reason about the effect of actions on belief, mutual belief, and intentions. Must be able to reason for any act and proposition P about whether or not [act]Believe{A, P) holds, and Believe(A, (act]P).

. Language production is not a faculty that can, in general, be ·solated from the general reasoning processes of a system. Natural communication requires knowing about the plans and goals of a speaker with respect to the entire task, and the ability to plan goals having to do with the communication process itself as weil as the domain. Therelore, it is impossible to take some existing system, add a user model, tack on a naturallanguage font end and back end, and expect it to engage in natural communication. The need for communication must be in the mind of the designer from the beginning, with domain and communication reasoning incorporated as a consistent whole.

USER MODELLING, COMMON-SENSE REASONING & the BELIEF-DESIRE-INTENSION PARADIGM

Kurt Konolige

User modelling is important wherever an Al system must interact with human agents. I say here "human agents", but this is not necessarily meant to exclude other types of agents; as computer system become more complex, the same principles used for efficient communication with people will hopefully apply to artificial agents. lndeed, in Methodologies below I note that analyzing the communication requirements of artificial agents may Iead to insights about communication in general.

I think it would be an understatement to say that current Al systems which incorporate user models have a long way to go. This is not because too little attention has been paid to the problem, but simply because the problern encompasses · a significant part of current Al research. There may be very restricted situations in which a crude parameter model (for example, a verbosity switch) is all that is necessary; but for the more open-ended dialogues that normally Iake place in question­ answering, explanation, and tutoring (to name a Iew application areas), a more accurate model ol the user's cognitive state is required. I would like to give a personal view of some of the major lines of research that are being pursued or should be pursued to achieve a realistic user model.

Methodologies

At present, most models of cognition in Al are variations of a 801 (belief-desire-intention) paradigm. An agent has beliefs about the world, and desires some states of the world more than others. Rational agents form intentions or plans to affect the state of the world to fulfill their desires, given the current state of their beliefs. This picture is a kind of commonsense psychology, and seems to be implicit in the way we use words like 'belief,' 'desire,' 'plans,' etc. Hardly any work has been done on a general theory relating these cognitive components. Still, the BOl paradigm is a useful general framework for constructing user models for particular applications. ln many cases, it is possible to simplify the model considerably: fcr example, in question-answering on a database it is assumed that the user has a goal of extracting information, and the problern of forming intentions from conflicting desires does not arise.

While the BOl paradigm can provide an overall hatrack for organizing cognitive models, it does not teil us what particular hats we should pul on it. Agents' beliefs, for example, can be quite complicated, incorporating complex commonsense reasoning about space, time, physical systems, and so on, as weil as particular beliefs about the domain at hand. How do we go about developing such theories? This might be called the Knowledge Problem for user modelfing. There are two sources for such theories. One is the Cognitive Science path, in which attention to protocols of subjects can yield interesting insight into cognitive processes acting in complex environments. The other is in Al planning systems: artificial agents whose cognitive structure is designed to solve a particular task. The former might be described as theory-poor but data-rich: the subjects actually do act intelligently in the domain, but the actual cognitive structures they employ are not accessible. The latt~r are theory-rich but data-poor: the design of the agent is useful as a theory of reasoning in the domain, but the agent may not actually act as intelligently as desired. So it would seem that both approaches are desirable .. for example, analyzing the way people use language yields data on desirable properlies of a language-using system, while studying the requirements for efficient communication between agents can Iead to a simplified model of communication that helps organize linguistic phenomena.

The Knowledge Problem

• At the very minimum, a useful user model for open-ended dialogues should include the following:
• Domain-dependent knowledge. This is the type of reasoning most ollen capture by expert systems, which are good at a very specialized type of problern in a narrowly defined setting.
• Theories of the commonsense world. This type of knowledge is tacitly assumed in all human communication. lt includes areas such as:
• lntentionality and beliefs of other agents.
• Common-sense theories of time, space, and physical processes.
• Knowledge of the Interaction process. This includes principles of efficient communication, such as "new Information comes first,'' or "use the most specific applicable term."

There is much significant work being done in Al on theories of this sort, e.g., work on qualitative reasoning, modelfing space and time, naive physics, logics of knowledge and belief, communication act theory, and so on.

The lnference Problem

The most important inference problern for user modelfing is the following:

Given the observed behavior of the user, find the appropriate state of the model that accounts for the behavior.

Note that this is a very different problern from performing inferences using one of the commonsense theories just mentioned. ln general, the latter is a deduction problem: find the consequences of a given theory. The inference problern for user modelfing is inductive: from a pattern of behavior, induce the correct structure that produces the behavior. Much of the work in script- or frame-based systems addresses this problem. However I think it is a much more difficult inference problem, and deserving of much moreintensive research.

USER MODELLING & PLAN RECOGNITION

N.S. Sridharan

A number of interesting and important questions have been raised for this panel. I wish to survey a small set of different tasks and show the diversity of responses possible depending on the characteristics of the task. I conclude by discussing a set of task dimensions which forms a framewerk for understanding, in a broad manner, the connections between tasks and user models.

1. Discussion of different domains
2. I.A Automatie prompting and simple help on workstations

An implicit assumption is made that the user may wish to know some information relevant to the command to be issued; and that it wont hurt to display such information automatically. A sketchy finite state machine model can be used compute allowable actions, allowable operands and a canned help text can be put up on the screen. No detailed user modelling is used.

1.8 Tutoring introductory programming: (EIIiott Soloway, Yale)

There is an intimate connection between plan recognition and user modelling. ln fact, not viewing user-modelling as plan recognition has hampered progress in this field. The stud~mt m!-lst be seen as trying to follow a plan; a program that 1s ~emg constructed is a realization of a plan. What the student IS trymg to accomplish, his goal is important; goal recognition is an importa~t problem. An approach using bug catalogs or plan catalogs IS inherently limited. That kind of approach will not go beyond small and simple programs. Plan recognitiön. must be viewed as .a constructive task; plan revision approach 1s very 1mportant. lt Will be not enough to think of selecting a user plan from a finite set of pre·formed plans. (For approaches to student modelling in tutoring domains where strong assumptions can be made about the student 's goal at any stage, see the earlier section Student Models in Intelligent Tutoring Systems.)

I.C Sensorsignal interpretation: (C.F. Schmidt, Rutgers)

The problern arises in connection with an int~ractive sy~tem. t~ assist in interpreting multiple unreliable sensor s1gnals. An 1mphc1t user model is used to effect (offline) tailoring of the system. Only the pragmatic consequences of accepting a model of user needs to be represented. The model used need only be accurate enough to predict the right actions; that is, the model is viewed only in terms of it implications for the system.

I.D Pilot's assistant: (Dick Pew, BBN)

ln automating the cockpit display for a military aircraft, it is clear that the display function . must be customized to the. use~. However in this domain, plans of pilots can only be deflned IS vague te~ms; e.g. as phases of a mission, and a conditional set of responses plus model of goals to maintain their priorities. The pilot is operating in an extremely dynamic situation, and opera~es generally by adopting opportunistic behavior and .reacttve behavior. ln this domain it appears that the user model IS better structured in terms of attributes such as focus of attention, span attention, attention switching speed, memory limitations, speed of observation and assimilation.

I.E Natural Language Dialog systems (Candy Sidner and Jim Schmolze, BBN)

Question-answering systems often Iimit themselves to dealing with individual questions separately; whereas, dialog systems attempt to include the context of the dialog so far. An important aspect of such contexts is a model of the user's inte~tions, capabilities, beliefs, knowledge, and preferences. Format1on of such a model is viewed as an incremental process. Understanding · communication requires accessing/hypoth~sizing the intentions. This process is one of forming a hypothes1s, and thus is inherently error-prone. The process used should be robust enough to recognize errorful hypotheses and to take steps to rectify them.

I.F Office automation task: non-linguistic, non-communicative domain (Vic Lesser, U of Mass.)

The user is engaged in a task such as filling out a purchase order for equipment. The system is watehing over his shoulders, so to speak, and attempting to guide the user. The user may be in error; the system is watehing to predict and correct the steps taken by the user. Plan recognition relies mostly on domain-based heuristics; and is less dependent oh modelling the user, his beliefs or plans. This is because the goal is to get the task accomplished rather than to train or educate the user.

II. Framewerk for discussion of user modelling

Very simple user models suffice for a number of tasks. ln spelling correction a simple model of user errors, rather than plans, can be immediately helpful. ln detecting errors in novices' program, a model of errors made frequently by novices, not their plans, may be very useful.

The attempt here is to set up a framewerk for exploring the analogies portrayed below.

• User customization (.) Prediction, correction,modification
• User models <·> Plans incorporating beliefs, intentions and goals
• User modelling process <·> Plan recognition

The purpose of the framewerk to be developed is to answer questions like: how complex should the user model be? what characteristics of the task are relevant in deciding how to acquire and use such user models?

Dimensions to consider and evaluate in exploring this analogy

1. Richness of response space: Assuming that the model is to guide a suitable action from a repertoire of actions, the user model must be (just) rich enough to guide choice of response, but shoulp be minimal. II the potential responses are not diverse the model can and should be quite simple.
2. Static vs Dynamic customization: lt is useful to consider whether the programmer is customizing the program to the user, or whether the system is adapting itself. Static customization may lend itself to implicit user models.
3. Who bears responsibility? system or user?: How complicated can the system become? ln an interactive Situation, the user will be formulating a model of the system, while the system is modelling the user. II the system is simple enough, the user may be willing to take responsibility for his own actions, since he can more readily form a model of the system. II the system is taking the responsibility for overall behavior, then the system should model the user accurately. (Biack box vs Glass box issue)

1. Risk or penalty for being w rong: Plan recognition and user modelling is inherently error-prone, being an inductive task. One must judge the consequence of this. II the penalty for using a wrang hypothesis is high, one musteilher not attempt to model the user or be in a setting where interactive verification of such hypotheses is feasible.

111 . What is user modelling? What is in a user model? There are different considerations that affect how one views the process of user modelling:

• Static vs Dynamic models (built·in vs acquired)
• Focus on immediate actions vs focus on eventual goals
• Deterministic vs probabilistic models
• Predictive vs descriptive models

Similarly one can imagine a variety of ways in which the user model is set up:

• ParamHter models
• State machine model (compute allowable actions; has a sense of history)
• Recursive models (user's model of the system; system self-model)
• Plan-based models (has a sense of goal). Beliefs, intentions and knowledge attributable to the user. Often these can be integrated in the form of a plan plus a context in which these plan are likely to be executed. Preferences, which allows the user to make choices.
• When should user modelling not be done?

There are a number of tasks where attempting to formulate a user model dynamically is inappropriate. Firstly, in some situations, users are evolving. One must realize that human beings are very adaptable and can evolve more rapidly than the systems. Secondly, in some Situations, users are not plan following, Hspecially if a user does not know what he wants. lt is futile to model fickle human beings. Thirdly, in some Situations the user may not know what the system can do. ln such Situations, a user lacking knowledge may only get confused if the system beings to alter its behavior. Fourthly, in some applications the boundaries ol responsibility between system and user may bE' shilting. For example, the user may wish to issue standing erde~ , thus causing the system to do some things routinely; or the user may wish to take away from the system certain tasks because the system too slow or unreliable. A highly interactive operating system, such as the TOPS·20, makes it quite dilficult to write script files lor automating certain lunctions - scripts cannot adequately substituteforahuman user.

V. Summary:

There is a number of problems where simple models werk reasonably weil; there are a number ol probleins where plan recognition is a challenge and can be made to werk weil if the assumptions about the user and task domain hold; there are task domains where the penalty lor being wrang is high and the acceptable complexity ol the model is low, that other techniques should be prolitably be pursued.

VI. Gonclusion:

Plan recognition, especially a constructive process that incorporates techniques for plan revision in addition to plan hypothesizing, see Schmidt, Sridharan & Goodson (1978), is a fundamental problern lor Artificial lntelligence. This problern deserves the same kind ol eifert that has been devoted to plan generation. The plan recognition problern is full of interesting challenges and intellectual surprises. lt ought to be pursued earnestly and serlously. Yet, in thinking of applications, one must be careful to use a suitable framewerk to decide what kind ol user. models and what type of user modelling process are best to adopt.

SUGGESTED READINGS

Appell, D. E. Planning English Sentences. Gambridge: Garnbridge University Press, 1985 (forthcoming).

Brown, J.S. & Burton, R.R. (1978). Diagnostic models for procedural bugs in basic mathematical skills, in Gognitive Science, 2. pp155-192.

Burton, R.R. & J.S. Brown (1982). An investigation of coinputer coaching for informal learning activities. ln Intelligent Tutoring Systems, eds D. Sleeman & J.S. Brown. Ac&demic: London, pp79-98.

Garbonell, J.R. (1970). Al in GAI: An Artificial lntelligence Approach to Gomputer-aided lnstruction. IEEE Transactions on Man-Machine Systems, 11, pp.190-202.

Gehen, P., and G. R. Perrault, "Elements of a Plan-Based Theory of Speech Acts." Gognitive Science, 3 (1979), pp. 177--212.

Konolige, K. Experimental Robot Psychology. Proceedings AGA/, (1985).

Mark, W., "Representation and lnference in the Gonsul System" ,IJGAI-81

Pollack, M., and J. Hirschberg "User Participation in the Reasoning Processes of Expert Systems." Proceedings of the National Conference on Artificallntelligence, (1982), pp. 358--361.

Reddy, M. J., "The Gonduit Metaphor---A Gase of Frame Conflict in Our Language about Language." ln Ortony (ed.), Metaphor and Thought, Gambridge: Garnbridge University Press1 (1979).

Rich, E. "Users are individuals: individualizing user models". IJMMS, 1983, volume 18, pp. 199-214.

Sleeman, D and J. S. Brown, lntroduction: Intelligent Tutering Systems. ln Intelligent Tutoring Systems D. Sleeman & J.S. Brown (eds) . Academic press: London. (1982) pp1·11.

Sleeman, D. UMFE: A User Modelling Front End system./JMMS, in press.

Schmidt, G.F., Sridharan, N.S., Goodson, J.L., The Plan Recognition Problem: An lntersection of Psychology and Artificial lntelligence, Artificiallntelligence, 11 (1978) pp45-83.

Swartout, W. , "The Gist Behavior Explainer," AAAI-8:J.

Swartout, W., "XPLAIN: a system for creating and explaining expert consulting programs," Artificiallntelligence, 21 (1983) pp. 285-325.