REQUIREMENTS FOR BELIEF MODELS IN COOPERATIVE DIALOGUE

Models of rationality typically rely on underlying logics that allow s imulated agents to entertain beliefs about one another to any depth of nesting. Such models seem to be overly complex when used for belief m odelling in environments in which cooperation between agents can be as sumed, i.e., most HCI contexts. We examine some existing dialogue syst ems and find that deeply-nested beliefs are seldom supported, and that where present they appear to be unnecessary except in some situations involving deception. Use of nested beliefs is associated with nested reasoning (i.e., reasoning about other agents' reasoning). We argue th at for cooperative dialogues, representations of individual nested bel iefs of the third level (i.e., what A thinks B thinks A thinks B think s) and beyond are in principle unnecessary unless directly available f rom the environment, because the corresponding nested reasoning is red undant. Since cooperation sometimes requires that agents reason about what is mutually believed, we propose a representation in which the se cond and all subsequent nesting levels are merged into a single catego ry. In situations affording individual deeply-nested beliefs, such a r epresentation restricts agents to human-like referring and repair stra tegies, where an unrestricted agent might make an unrealistic and perp lexing utterance.