Statistical Voice Dialog System and Method

• US 20150228275A1
• Filed: 02/10/2014
• Published: 08/13/2015
• Est. Priority Date: 02/10/2014
• Status: Active Grant

• Current Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Original Assignee: Mitsubishi Electric Research Laboratories, Inc. (Mitsubishi Electric Corporation)
• Inventors: Watanabe, Shinji, Hershey, John R.
• Granted Patent: US 9,837,075 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G10L 15/083 Recognition networks G10L15/142, G10L15/16 take precedence
• CPC Class Codes: G10L 15/142 Hidden Markov Models [HMMs]
• CPC Class Codes: G10L 15/1822 Parsing for meaning understanding
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

A method for processing a voice command using a statistical dialog model determines a belief state as a probability distribution over states organized in a hierarchy with a parent-child relationship of nodes representing the states. The belief state includes the hierarchy of state variables defining probabilities of each state to correspond to the voice command and a probability of a state of a child node in the hierarchy is conditioned on a probability of a state of a corresponding parent node. A system action is selected based on the belief state.

Claims

1. A method for processing a voice command using a statistical dialog model, comprising:

determining, in response to receiving the voice command, a belief state as a probability distribution over states organized in a hierarchy with a parent-child relationship of nodes representing the states, such that the belief state includes the hierarchy of state variables defining probabilities of each state to correspond to the voice command, wherein a probability of a state of a child node in the hierarchy is a conditioned on a probability of a state of a corresponding parent node; and

selecting a system action based on the belief state, wherein steps of the method are performed by a processor.

2. The method of claim 1, wherein the determining comprises:

determining the belief state as a product of conditional probabilities of each child node given its parent nodes.

3. The method of claim 1, wherein the hierarchy of state variables includes non-zero probabilities for at least two child nodes having different parents.

4. The method of claim 1, wherein the hierarchy is a multi-level hierarchical structure including a directed acyclic graph or a tree.

5. The method of claim 1, wherein the selecting is performed according to an optimization of a partially observable Markov decision process.

6. The method of claim 1, wherein the selecting comprises:

selecting a subset of the state variables from different levels of the hierarchy; and

selecting the system action by optimizing an objective function applied to the subset of the state variables.

7. The method of claim 6, further comprising:

selecting N-best-scored child nodes from one or more levels of the hierarchy, wherein the N-best-scored child nodes have highest probabilities among child nodes of their level, wherein N is an integer; and

determining the subset of the state variables according to the probabilities of the N-best scored child nodes.

8. The method of claim 7, wherein N varies among the levels of the hierarchy.

9. The method of claim 1, further comprising:

selecting a subset of state variables with highest values among state variables of corresponding levels;

selecting a set of candidate actions associated with states of the subset of state variables; and

selecting the system action from the set of candidate actions by optimizing an objective function applied to the subset of the state variables.

10. A statistical voice dialog system for processing a voice command, comprising:

a storage unit for storing a hierarchy of states with a parent-child relationship of nodes representing the states, such that a probability of a state of a child node to correspond to the voice command is conditioned on a probability of a state of a parent node to correspond to the voice command;

an update unit for updating a probability distribution over the states using a product of conditional probabilities of each child node given its parent nodes to produce a belief state including the hierarchy of state variables defining probabilities of each state to correspond to the voice command; and

a command unit for selecting a system action based on the belief state.

11. The system of claim 10, wherein the update unit updates the probability distribution based on a previous probability distribution, a previous system action, a priori state transition probabilities, and a current probability distribution inferred from the voice command, and wherein the command unit selects the system action based on the updated probability distribution such that an objective function is optimized.

12. The system of claim 10, wherein states are associated with leaf nodes or with intermediate nodes, such that a path from a root node to a leaf node through an intermediate node defines at least one state, wherein prompt actions are associated with the state of the intermediate nodes, and command actions are associated with the states of the leaf nodes.

13. The system of claim 10, further comprising:

an error determining unit for determining an error in recognizing the voice command and/or an error in the probability distribution; and

a re-input unit for generating a response requesting re-input of the voice command.

14. The system of claim 10, wherein the update unit update the probability distribution for each level in a top-down manner, such that probabilities of nodes of a level are determined after the probabilities of nodes of a previous level.

15. The system of claim 10, wherein the command unit selects a subset of state variables with highest values among state variables of corresponding levels, selects a set of candidate actions corresponding to the subset of state variables, and selects the system action from the set of candidate actions by optimizing an objective function applied to the subset of the state variables.

16. The system of claim 15, wherein the objective function minimizes a cost with respect to the belief state, wherein the cost is a function of a mismatch between the selected system action and an intention of the voice command as modeled by the system.

17. A method for processing a voice command using a statistical dialog model, comprising:

determining, in response to receiving the voice command, a belief state as a probability distribution over states organized in a hierarchy with a parent-child relationship of nodes representing the states, such that the belief state includes the hierarchy of state variables defining probabilities of each state to correspond to the voice command, wherein a probability of a state of a child node in the hierarchy is conditioned on a probability of a state of a corresponding parent node, and wherein the states of the hierarchy are states of Partially Observable Markov Decision Process (POMDP);

selecting a subset of states from multiple levels of the hierarchy with highest value of state variables;

selecting a set of candidate actions corresponding to the subset of states; and

selecting the system action from the set of candidate actions by solving the POMDP.