Belief tracking and action selection in spoken dialog systems

• US 8,676,583 B2
• Filed: 08/30/2011
• Issued: 03/18/2014
• Est. Priority Date: 08/30/2010
• Status: Expired due to Fees

• Current Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Original Assignee: Honda Motor Co., Ltd. (Honda Motor Company)
• Inventors: Gupta, Rakesh, Ramachandran, Deepak, Raux, Antoine, Mehta, Neville, Krawczyk, Stefan, Hoffman, Matthew
• Primary Examiner(s): He, Jialong
• Assistant Examiner(s): Shan, Jie
• Application Number:
  US13/221,155
• Publication Number:
  US 20120053945A1
• Time in Patent Office: 931 Days
• Field of Search: 704/256
• US Class Current: 704/256
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

An action is performed in a spoken dialog system in response to a user'"'"'s spoken utterance. A policy which maps belief states of user intent to actions is retrieved or created. A belief state is determined based on the spoken utterance, and an action is selected based on the determined belief state and the policy. The action is performed, and in one embodiment, involves requesting clarification of the spoken utterance from the user. Creating a policy may involve simulating user inputs and spoken dialog system interactions, and modifying policy parameters iteratively until a policy threshold is satisfied. In one embodiment, a belief state is determined by converting the spoken utterance into text, assigning the text to one or more dialog slots associated with nodes in a probabilistic ontology tree (POT), and determining a joint probability based on probability distribution tables in the POT and on the dialog slot assignments.

Claims

1. A computer-based method of performing an action in a spoken dialog system comprising:

retrieving a policy mapping belief states of user intents to actions;

receiving a spoken utterance from a user;

determining a belief state of the user'"'"'s intent based on the spoken utterance using a probabilistic ontology tree (POT) comprising a plurality of unobserved nodes each representing a domain concept and comprising a probability distribution table including probabilities that the concept represented by the unobserved node takes each of a plurality of values, wherein each unobserved node is associated with a dialog slot to which the spoken utterance is assigned based on the relatedness between the spoken utterance and the domain concept represented by the unobserved node;

selecting an action to take based on the determined belief state and the retrieved policy; and

performing the selected action.

2. The method of claim 1, wherein the selected action comprises an action requiring additional feedback from the user.

3. The method of claim 2, wherein the selected action comprises a request for clarification of all or part of the spoken utterance from the user.

4. The method of claim 2, wherein the selected action comprises a request for confirmation of all or part of the spoken utterance from the user.

5. The method of claim 1, wherein the selected action comprises informing the user of information requested by the user.

6. The method of claim 1, wherein retrieving a policy comprises creating a policy.

7. The method of claim 6, wherein creating a policy comprises:

creating policy parameters;

creating a simulated user input with a pre-determined goal;

determining a simulated belief state of user intent based on the created simulated user input;

performing an action based on the simulated belief state and the policy parameters; and

outputting the policy parameters if the performed action satisfies the pre-determined goal.

8. The method of claim 7, further comprising:

modifying the simulated user input based on the performed action if the performed action does not satisfy the pre-determined goal;

determining a second simulated belief state of user intent based on the modified simulated user input;

performing a second action based on the second simulated belief state and the policy parameters; and

outputting the policy parameters if the performed second action satisfies the pre-determined goal.

9. The method of claim 8, further comprising:

determining a partially observable Markov decision process (POMDP) reward score based on the performed actions; and

responsive to a determination that a policy threshold is satisfied based on the POMDP reward score, outputting the policy parameters.

10. The method of claim 9, further comprising:

responsive to a determination that a policy threshold is not satisfied based on the POMDP reward score, modifying the policy parameters based on one or more of;

the simulated belief state, the second simulated belief state, the performed action, the second performed action, and the POMDP reward score;

creating a second simulated user input with a pre-determined goal;

determining a third simulated belief state of user intent based on the second simulated user input; and

performing a third action based on the third simulated belief state and the modified policy parameters.

11. The method of claim 1, wherein the determined belief state comprises belief features of the user'"'"'s intent.

12. The method of claim 1, wherein determining a belief state of the user'"'"'s intent based on the spoken utterance comprises:

creating an observed node in the POT for each dialog slot associated with an unobserved node to which the spoken utterance is assigned, each observed node comprising a probability distribution table including probabilities that the intended node value of the unobserved node is correctly and incorrectly identified; and

determining a belief state of user intent based on the joint probability of the unobserved node probability distribution tables and the observed node probability distribution tables.

13. The method of claim 12, wherein the spoken utterance is converted into text using automatic speech recognition, and wherein the spoken utterance is assigned to one or more dialog slots by determining a confidence score for each assignment based on the likelihood that the text was correctly assigned to a dialog slot.

14. The method of claim 13, wherein the probability distribution tables of the observed nodes are determined based on the confidence score and the domain concept associated with the observed nodes.

15. The method of claim 12, wherein the joint probability of the unobserved node probability distribution tables and the observed node probability distribution tables is determined using a message-passing algorithm in which child node passes a message to the child node'"'"'s parent node, and wherein each message comprises the joint probability distributions of the child node and all descendant nodes of the child node.

16. A spoken dialog system for performing an action comprising:

a policy module for retrieving a policy mapping belief states of user intents to actions;

a microphone for receiving a spoken utterance from a user;

a belief tracking module for determining a belief state of the user'"'"'s intent based on the spoken utterance using a probabilistic ontology tree (POT) comprising a plurality of unobserved nodes each representing a domain concept and comprising a probability distribution table including probabilities that the concept represented by the unobserved node takes each of a plurality of values, wherein each unobserved node is associated with a dialog slot to which the spoken utterance is assigned based on the relatedness between the spoken utterance and the domain concept represented by the unobserved node;

a policy interface for selecting an action to take based on the determined belief state and the retrieved policy; and

a system interface for performing the selected action.

17. A non-transitory computer-readable storage medium having computer-executable code for performing an action comprising:

a policy module configured to retrieve a policy mapping belief states of user intents to actions;

a microphone module configured to receive a spoken utterance from a user;

a belief tracking module configured to determine a belief state of the user'"'"'s intent based on the spoken utterance using a probabilistic ontology tree (POT) comprising a plurality of unobserved nodes each representing a domain concept and comprising a probability distribution table including probabilities that the concept represented by the unobserved node takes each of a plurality of values, wherein each unobserved node is associated with a dialog slot to which the spoken utterance is assigned based on the relatedness between the spoken utterance and the domain concept represented by the unobserved node;

a policy interface module configured to select an action to take based on the determined belief state and the retrieved policy; and

a system interface module configured to perform the selected action.

18. A computer-based method for determining a belief state of a user'"'"'s intent comprising:

receiving a spoken utterance from the user;

converting the spoken utterance into text;

retrieving a probabilistic ontology tree (POT), the POT comprising a plurality of unobserved nodes, each unobserved node representing a domain concept and comprising a probability distribution table including probabilities that the concept represented by the unobserved node takes each of a plurality of values, wherein each unobserved node is associated with a dialog slot;

assigning the text to one or more dialog slots associated with unobserved nodes based on the relatedness between the text and the domain concepts represented by the unobserved nodes;

creating an observed node in the POT for each dialog slot to which text is assigned, the observed nodes comprising a probability distribution table including probabilities that the intended node value of the unobserved node is correctly and incorrectly identified; and

determining a belief state of user intent based on the joint probability of the unobserved node probability distribution tables and the observed node probability distribution tables.