Introduction
In the realm of intelligent advisors—systems such as sales assistants or personalized news filters—the primary objective is to provide information that assists a user (the "evaluator") in making informed judgments. This paper introduces a unified framework for "Evaluation-Oriented Information Provision" (EOIP). By integrating Multi-Attribute Utility Theory (MAUT) with Bayesian networks, the authors address a critical challenge in advisor systems: managing uncertainty regarding a user's preferences and prior expectations.
The EOIP Framework
Historically, advisor systems relied on ad-hoc rules or rigid heuristics. The proposed framework formalizes the interaction by decomposing objects into structured "value trees."
Multi-Attribute Utility Theory (MAUT)
Under MAUT, an object (e.g., a car) is represented as a hierarchy of dimensions and attributes. Each leaf node represents a concrete attribute, such as mileage or the presence of airbags. These attributes are assigned value functions to determine their utility and importance weights to reflect their significance to the user.
Modeling Uncertainty with Bayesian Networks
A central innovation of this work is the use of Bayesian networks to represent the system's uncertainty about the user. Instead of assuming fixed preferences, the system treats importance weights and user expectations as probability distributions.
Predicting Evaluation Shifts
Rather than simply predicting an absolute evaluation (e.g., "the user likes this car"), the system focuses on the "evaluation shift." This represents the change in a user's internal assessment triggered by a specific piece of information. If a system knows a user expects a car to have high mileage, informing them that the mileage is actually low creates a significant positive shift, making that information highly relevant for the dialog.
Upward Propagation and Elicitation
The framework supports two-way information flow:
- Interpretation: When a user reacts (e.g., "That's good"), the system uses Bayesian upward propagation to update its model of what the user values.
- Elicitation: The system calculates which questions will most effectively reduce its uncertainty about the user's "target variables," such as their specific weighting of safety versus reliability.
Practical Implementation: The PRACMA System
The authors implemented this framework in PRACMA, a dialog system for the used-car domain. PRACMA demonstrates "mixed-initiative" dialog, where the system can follow a structured sales strategy while remaining flexible enough to handle unexpected user input. By treating user preferences as distributions rather than points, PRACMA avoids the overconfident, often incorrect recommendations typical of earlier heuristic-based systems.
Conclusion
By categorizing EOIP into five core tasks—predicting overall and partial evaluations, interpreting evidence, eliciting evidence, and selecting dialog moves—the authors provide a principled roadmap for future advisor systems. The shift from absolute evaluations to probabilistic evaluation shifts allows for more nuanced, helpful, and human-like information provision.
Summary updated