Plan recognition is an important aspect of dialogue processing, because it provides access to the speaker's intentions. Goal analysis is plan recognition that focuses on inferring the speaker's goals. This dissertation describes a goal analyzer called PAGAN that is part of the UNIX Consultant project.

This dissertation makes four main contributions. First, it provides a theory of how the merits of an explanation of an utterance may be judged. There are three criteria for making such judgements: the applicability of the explanation to the system's needs; the grounding of the explanation in what is already known of the speaker and of the dialogue; and the completeness of the explanation's coverage of the speaker's goals.

The second main contribution of this work is the introduction of a representation for plans and goals that solves two problems inherent in many approaches to plan representation. First, there is usually an inadequate distinction between the effects of an action and the effect that the action was intended to produce. Secondly, the traditional notion of a precondition fails to distinguish between heuristic planning knowledge and knowledge of the defining properties of actions. The solution to these problems, called a planfor, is a relation between a type of goal and a sequence of hypothetical actions that constitutes a possible method of achieving a goal of that type.

The third contribution of this work is a detailed analysis of ambiguity. The sources of ambiguity include ambiguities at the sentence and utterance levels, ambiguity arising from the existence of multiple explanatory plan schemas, ambiguity arising when it is possible to place a mentioned concept into more than one competing category, and ambiguity arising when it is unclear which concept is the topic of a particular relation (such as a question). This dissertation suggests a unified approach to handling these types of ambiguities.

The fourth contribution of this work is the introduction of a method for determining how much processing of an utterance should be done. It is based on an assessment of the completeness of the explanation, and on an assessment of the practicability of continuing the processing.




Download Full History