COMPUTER-BASED REPRESENTATIONS AND REASONING METHODS FOR ENGAGING USERS IN GOAL-ORIENTED CONVERSATIONS

The use of a goal-understanding abstraction hierarchy in conjunction with Bayesian inference, decision-theoretic analyses of conversational and observational actions is disclosed to provide machinery for incremental refinement of an understanding about a user's goals through conversation with users. A computer-implemented method receives multiple classes of information regarding a user goal including visual and linguistic clues at a specific level of the abstraction hierarchy, to assess the goal. The method then determines a with a value-of-information analysis the utility of acquiring additional information via making additional observations or by explicitly querying the user versus making a decision to change the level of precision of the analysis of a user's goals. Throughout the analysis, a probability distribution is inferred about the goals of a user. This probability distribution is used in conjunction with a representation of utility of different outcomes to identify informational and navigational actions with the greatest expected utility. In one embodiment, the probability of the leading goal is inferred and used to drive decision making, for example, in assuming the relevance of particular sub-goals of the current goal, where the sub-goals are in a succeeding level of the hierarchy. The probability can be determined in one embodiment by a Bayesian network. If the highest probability sub-goal exceeds a progression threshold, which can be determined by an approximate decision analysis, then this sub-goal is proceeded to in one embodiment—that is, the current level is advanced to the succeeding level, and information gathering is initiated at this new level.