DIALOGUE SYSTEM, A DIALOGUE METHOD AND A METHOD OF ADAPTING A DIALOGUE SYSTEM

• US 20180233143A1
• Filed: 08/28/2017
• Published: 08/16/2018
• Est. Priority Date: 02/13/2017
• Status: Active Grant

• Current Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Original Assignee: Kabushiki Kaisha Toshiba (Toshiba Corporation)
• Inventors: PAPANGELIS, Alexandros, STYLIANOU, loannis
• Granted Patent: US 10,446,148 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 16/3329 Natural language query formulation or dialogue systems
• CPC Class Codes: G06F 40/237 Lexical tools
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

A dialogue system comprising:

• • an input for receiving data relating to a speech or text signal originating from a user;
  • an output for outputting information specified by an action; and
  • a processor configured to:
    • update one or more system states based on the input data using one or more state tracker models, wherein the one or more system states comprise probability values associated with each of a plurality of possible values for each of a plurality of categories, wherein a category corresponds to a subject that the speech or text signal may relate to and can take on one or more values from a set of values;
    • determine an action function and determine an action function input by inputting information generated using the system state and a set of stored information into a policy model, the set of stored information comprising a plurality of action functions;
    • output information specified by the determined action function and the determined action function input at the output.

Claims

1. A dialogue system comprising:

an input for receiving data relating to a speech or text signal originating from a user;

an output for outputting information specified by an action; and

a processor configured to;

update one or more system states based on the input data using one or more state tracker models, wherein the one or more system states comprise probability values associated with each of a plurality of possible values for each of a plurality of categories, wherein a category corresponds to a subject that the speech or text signal may relate to and can take on one or more values from a set of values;

determine an action function and determine an action function input by inputting information generated using the system state and a set of stored information into a policy model, the set of stored information comprising a plurality of action functions;

output information specified by the determined action function and the determined action function input at the output.

2. The dialogue system according to claim 1, wherein the plurality of categories are from a plurality of different pre-defined domain-specific ontologies, wherein a pre-defined domain corresponds to a particular dialogue topic and wherein the pre-defined domain-specific ontology comprises the plurality of categories relating to the particular dialogue topic, and wherein the plurality of action functions are domain independent.

3. The dialogue system according to claim 2, wherein the inputted information generated using the system state or the set of stored information comprises a set of possible action function inputs, and wherein determining an action function and determining an action function input comprises:

for each possible action function input, generating a vector of values, wherein each value in each vector corresponds to an action function and is an estimate of the dialogue performance if the action function and action function input are used to generate the output information, the values being generated from stored parameters;

wherein the determined action function and action function input corresponds to the action function corresponding to the highest value from all of the vectors.

4. The dialogue system according to claim 3, wherein the policy model is configured to operate with ontology-independent parameters as input and the processor is further configured to:

parameterize the information generated using the system state in terms of one or more ontology independent parameters.

5. The dialogue system according to claim 4, wherein each action function input is parameterised in terms of one or more ontology independent parameters and the parameters used to generate the vector of values for the action function input, wherein each vector of values corresponds to a domain specific action function input.

6. The dialogue system according to claim 3, wherein the vectors are generated using a neural network.

7. The dialogue system according to claim 1, wherein the input is for receiving a speech signal and wherein the output is an output for outputting a speech signal and wherein the processor is further configured to:

generate an automatic speech recognition hypothesis with associated probability from the input speech signal

generate text based on the action;

convert the text to speech for outputting at the output.

8. The dialogue system according to claim 1, wherein the inputted information generated using the system state or the set of stored information comprises a set of possible action function inputs, wherein the action function inputs comprise categories, wherein the processor is further configured to:

identify one or more relevant categories;

generate a reduced system state, comprising the probability values associated with one or more of the plurality of possible values for each of the relevant categories, wherein at least part of the reduced system state is inputted into the model.

9. The dialogue system according to claim 8, wherein identifying one or more relevant categories comprises inputting at least part of the updated system state information into an identifier model, which obtains a probability that each category is relevant based on the updated system state and identifies the relevant categories as those with probabilities higher than a threshold value, wherein the relevant categories comprise categories from different pre-defined domain-specific ontologies and wherein the reduced system state does not comprise probability values associated with any of the plurality of possible values for any of the categories which were not identified as relevant.

10. The dialogue system according to claim 8, wherein identifying one or more relevant categories comprises inputting at least part of the updated system state information into a dialogue topic tracker model, which obtains a probability that each dialogue topic is relevant based on the updated system state and identifies the most relevant dialogue topic, wherein the relevant categories comprise categories from the pre-defined domain-specific ontology corresponding to the most relevant dialogue topic and wherein the reduced system state does not comprise probability values associated with any of the plurality of possible values for any of the categories which were not identified as relevant.

11. The dialogue system according to claim 1, wherein the categories are slots.

12. The dialogue system according to claim 1, wherein updating one or more system states comprises updating a plurality of system states based on the input data using a plurality of state tracker models, wherein each system state comprises probability values associated with each of a plurality of possible values for each of a plurality of categories associated with a pre-defined domain.

13. A dialogue method comprising:

receiving data relating to a speech or text signal originating from a user;

updating one or more system states based on the input data using one or more state tracker models, wherein the one or more system states comprise probability values associated with each of a plurality of possible values for each of a plurality of categories, wherein a category corresponds to a subject that the speech or text signal may relate to and can take on one or more values from a set of values;

determining an action function and determining an action function input by inputting information generated using the system state and a set of stored information into a policy model, the set of stored information comprising a plurality of action functions;

outputting information specified by the determined action function and the determined action function input at the output.

20. A carrier medium comprising computer readable code configured to cause a computer to perform the method of claim 13.

14. A method of adapting a dialogue system, by using the system repeatedly to run through dialogues with a human or a simulated human and providing a performance indicator for each dialogue, and wherein the method further comprises:

receiving data relating to a speech or text signal in a dialogue;

updating a system state based on the input data using a state tracker model, wherein the system state comprises probability values associated with each of a plurality of possible values for each of a plurality of categories, wherein a category corresponds to a subject that the speech or text signal may relate to and can take on one or more values from a set of values;

determining an action function and determining an action function input by inputting information generated using the system state and a set of stored information into a policy model, the set of stored information comprising a plurality of action functions;

outputting information specified by the determined action function and the determined action function input at the output;

adapting the policy model to increase the performance indicator.

15. The method according to claim 14, wherein the inputted information generated using the system state or the set of stored information comprises a set of possible action function inputs, and wherein determining an action function and determining an action function input comprises:

for each possible action function input generating a vector of values, wherein each value in each vector corresponds to an action function and is an estimate of the dialogue performance if the action function and action function input are used to generate the output information, the values being generated from stored parameters;

wherein the determined action function and action function input corresponds to the action function corresponding to the highest value from all of the vectors; and

wherein adapting the policy model comprises updating the stored parameters based on the performance indicator.

16. The method according to claim 14, wherein the inputted information generated using the system state or the set of stored information comprises a set of possible action function inputs, wherein the action function inputs comprise categories, and further comprising:

identifying one or more relevant categories;

wherein the system state is a reduced system state, comprising the probability values associated with one or more of the plurality of possible values for each of the relevant categories.

17. The method according to claim 16, wherein the one or more relevant categories are identified based on at least part of an input system state information using an identifier model.

18. The method according to claim 14, further comprising:

for each category present in the dialogues, determining a cumulative moving average of the ratio of the performance indicator achieved to the maximum possible performance indicator as an estimate of category importance, wherein the importance is used as one of the second set of parameters.

19. The method according to claim 14, further comprising:

for each category in the dialogues, determining a cumulative moving average of the relative position in the dialogue as an estimate of category priority, wherein the priority is used as one of the second set of parameters.