Uncertainty plays a central role in the handling of misunderstanding in dialog. People engaged in conversation typically take a sequence of actions to establish and maintain mutual understanding -- a process referred to as grounding. We explore representations and control strategies for grounding utterances founded on performing explicit probabilistic inference about failures in communication. The methods are informed by psychological studies and founded on principles of decision making under uncertainty. We delineate four distinct levels of analysis for representing uncertainty and describe a computational framework for guiding action in an automated conversational system. We demonstrate how the framework captures grounding behavior by facilitating collaborative resolution of uncertainty as implemented in a spoken interactive dialog prototype called the Bayesian Receptionist.