Hierarchical belief states for digital assistants

A method for operating a digital assistant, the method comprising:

• at an electronic device having one or more processors and memory:
• receiving a user utterance of a dialogue;
• parsing one or more text representations of the user utterance to determine a plurality of semantic interpretations for the user utterance, the plurality of semantic interpretations including more than two concepts or properties, wherein the parsing includes determining that a first concept or property and a second concept or property in the more than two concepts or properties have a joint semantic relationship;
• determining, based on the plurality of semantic interpretations, a belief state for the dialogue, the belief state comprising a plurality of dialogue slots representing the more than two concepts or properties, wherein each dialogue slot of the plurality of dialogue slots includes a respective marginal certainty for a respective concept or property represented by the respective dialogue slot, and wherein a first dialogue slot of the plurality of dialogue slots further includes a joint certainty for the first concept or property and the second concept or property based on the determined joint semantic relationship;
• determining a plurality of candidate policy actions from the determined belief state;
• selecting, based on the marginal certainty of each dialogue slot of the plurality of dialogue slots and the joint certainty, a policy action from the plurality of candidate policy actions;
• and
• performing the selected policy action, including outputting results of the policy action for presentation.

A non-transitory computer-readable storage medium storing one or more programs configured to be executed by one or more processors of an electronic device and memory, the one or more programs including instructions for:

• receiving a user utterance of a dialogue;
• parsing one or more text representations of the user utterance to determine a plurality of semantic interpretations for the user utterance, the plurality of semantic interpretations including more than two concepts or properties, wherein the parsing includes determining that a first concept or property and a second concept or property in the more than two concepts or properties have a joint semantic relationship;
• determining, based on the plurality of semantic interpretations, a belief state for the dialogue, the belief state comprising a plurality of dialogue slots representing the more than two concepts or properties, wherein each dialogue slot of the plurality of dialogue slots includes a respective marginal certainty for a respective concept or property represented by the respective dialogue slot, and wherein a first dialogue slot of the plurality of dialogue slots further includes a joint certainty for the first concept or property and the second concept or property based on the determined joint semantic relationship;
• determining a plurality of candidate policy actions from the determined belief state;
• selecting, based on the marginal certainty of each dialogue slot of the plurality of dialogue slots and the joint certainty, a policy action from the plurality of candidate policy actions;
• and
• performing the selected policy action, including outputting results of the policy action for presentation.