Adaptive Provision of Evaluation-Oriented Information: Tasks and Techniques

Evaluation-oriented information provision is a function performed by many systems that serve as personal assistants, advisors, or sales assistants. Five general tasks are distinguished which need to be addressed by such systems. For each task, techniques employed in a sample of systems are discussed, and it is shown how the lessons learned from these systems can be taken into account with a set of unified techniques that make use of well-understood concepts and principles from Multi-Attribute Utility Theory and Bayesian networks. These techniques are illustrated as realized in the dialog system PRACMA. During the past two decades, a number of AI systems have been developed whose overall task can be characterized as evaluation-oriented information provision: The user (to be called the evaluator, or E) has the goal of making evaluative judgments about one or more objects; the system (or information-provider, I) supplies E with information to help E make these judgments. Table 1 lists a representative sample of five such systems, which will be referred to as EOIPs.1 The number of such systems seems likely to grow in the near future, especially given the recent interest in personal assistants —some of which advise their users on evaluative judgments —and teleshopping, which should increase the demand for automated sales assistants. EOIP systems differ considerably in the techniques they employ for interaction with the user and for internal processing. For example, the communication with the INFORMATION FILTERING SYSTEM and the SALES ASSISTANT is realized with direct manipulation and hypertext techniques, whereas the other three systems use some form of natural language. There are also large differences in the theoretical frameworks and terminology in which EOIPs are presented. These differences impede exchange and consolidation of results. The present paper aims (a) to remedy this state of affairs by providing a uniThis research is being supported by the German Science Foundation (DFG) in its Special Collaborative Research Program on Artificial Intelligence and Knowledge-Based Systems (SFB 314), project N1, PRACMA. Not included are expert systems that perform evaluation tasks using evaluation criteria that have no necessary relationship to the criteria of the user (see, e.g., [Klein and Shortliffe, 1994]). Table 1: Overview of Five Representative Systems for Evaluation-Oriented Information Provision Example Example System Reference Evaluator Objects GRUNDY [Rich, 1979] Library user Library books INFORMATION [Sheth and Reader of Individual FILTERING Maes, 1993] network news news SYSTEMa articles CONSULTb [Elzer et al., University University 1994] student courses SALES [Popp and Potential buyer Products ASSISTANT Lödel, 1994] (e.g. of personal offered for computer) sale PRACMA [Jameson et Potential usedAvailable al., 1994] car buyer cars aNo system name was given in the cited paper. bThis name was introduced after the appearance of the cited paper. fied framework for analyzing the techniques used in EOIPs; and (b) to advance the state of the art by presenting some new techniques which should be generally applicable within EOIPs. Table 2 gives an overview of five general tasks which are at least potentially relevant to any EOIP. These will be discussed in turn in the five sections to follow. The new techniques will be presented in the context of the fifth of the reference systems, PRACMA. The excerpt from an example dialog in Table 3 both gives a sense of the nature of PRACMA’s dialogs and provides initial examples of the five tasks. 1 Task 1: Predict Overall Evaluations It is almost inevitable for an EOIP to try to predict how the user E would evaluate individual objects in the domain if he2 had complete information about them. For example, though it is clear in the used-car domain that the buying decision will ultimately be made by E , I needs to predict E’s overall evaluations in order to narrow the discussion to one or more For clarity, masculine and neuter pronouns will be used to refer to E and I, respectively. Table 2: Overview of Five General Tasks for an EvaluationOriented Information Provider 1. Predict Overall Evaluations: Anticipate how E would evaluate one or more domain objects, perhaps relative to one another, given complete knowledge about them. 2. Predict Partial Evaluations: Anticipate the impact that information about an attribute of an object would have on E’s evaluation of that object. 3. Interpret Evidence: Update the model of E’s evaluation criteria on the basis of evidence in E’s actions. 4. Elicit Evidence: Induce E to perform actions that will constitute evidence for the task “Interpret Evidence”. 5. Select Dialog Moves: Determine what type of dialog move to make (e.g., formulate recommendation; ask question about E’s criteria; allow E to act next).