Towards a Computational Theory of Grounding in Natural Language Conversation

: Theories of speech acts view utterances as actions which attempt to change the mental states of agents participating in a conversation. Recent work in computer science has tried to formalize speech acts in terms of the logic of action in AI planning systems. However most of these planning systems make simplifying assumptions about the world which are too strong to capture many features of conversation. One of these assumptions has been that the intent of an utterance is mutually understood by participants in a conversation, merely in virtue of its having been uttered in their presence. Clark and Marshall, 1981 have assumptions of attention, rationality, and understandability to accomplish this. Perrault, 1990 uses an assumption of observability. While these assumptions may be acceptable for processing written discourse without time constraints, they are not able to handle a large class of natural language utterances, including acknowledgements, and repairs. These phenomena have been studied in a descriptive fashion by sociologists and psychologists. I present ideas leading to a computational processing model of how agents come to reach a state of mutual understanding about intentions behind utterances. This involves a richer, hierarchical notion of speech acts, and models for tracking the state of knowledge in the conversation.