Systems and processes for operating a digital assistant using a hierarchical belief state are disclosed. In an example process, a user utterance of a dialog is received. A belief state for the dialog is determined. The belief state comprises a plurality of dialog slots. Each dialog slot of the plurality of dialog slots includes a respective marginal certainty for a concept or property represented by the respective dialog slot. A first dialog slot of the plurality of dialog slots further includes one or more joint certainties for one or more interpretations arising from the first dialog slot. Based on the marginal certainty of each dialog slot of the plurality of dialog slots and the one or more joint certainties of the first dialog slot, a policy action is selected from a plurality of candidate policy actions that correspond to the belief state. The selected policy action is performed.