Abductive Reasoning for Continual Dialogue Understanding

In this paper I present a continual context-sensitive abductive framework for understanding situated spoken natural dialogue. The framework builds up and refines a set of partial defeasible explanations of the spoken input, trying to infer the speaker’s intention. These partial explanations are conditioned on the eventual verification of the knowledge gaps they contain. This verification is done by executing test actions, thereby going beyond the initial context. The approach is illustrated by an example set in the context of human-robot interaction.