Dialog repair based on discrepancies between user model predictions and speech recognition results

• US 20070233497A1
• Filed: 03/30/2006
• Published: 10/04/2007
• Est. Priority Date: 03/30/2006
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Paek, Timothy, Chickering, David
• Granted Patent: US 8,244,545 B2
• Time in Patent Office: Days
• US Class Current: 704/270
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue
• CPC Class Codes: G10L 2015/228 of application context

Abstract

An architecture is presented that leverages discrepancies between user model predictions and speech recognition results by identifying discrepancies between the predictive data and the speech recognition data and repairing the data based in part on the discrepancy. User model predictions predict what goal or action speech application users are likely to pursue based in part on past user behavior. Speech recognition results indicate what goal speech application users are likely to have spoken based in part on words spoken under specific constraints. Discrepancies between the predictive data and the speech recognition data are identified and a dialog repair is engaged for repairing these discrepancies. By engaging in repairs when there is a discrepancy between the predictive results and the speech recognition results, and utilizing feedback obtained via interaction with a user, the architecture can learn about the reliability of both user model predictions and speech recognition results for future processing.

Claims

1. A system for repairing dialog data, comprising:

a discrepancy detection component that identifies discrepancy data between predictive dialog data output from a user model prediction component and recognized dialog data output from a speech recognition component; and

a dialog repair component that repairs the dialog data based in part on the discrepancy data.

2. The system of claim 1, wherein the discrepancy data is based on a measure of difference between the predictive dialog data and the recognized dialog data.

3. The system of claim 1, wherein the discrepancy data is based on a measure of reliability, the measure of reliability is at least one of probability, confidence score, and potential functions.

4. The system of claim 1, wherein at least one of the predicted dialog data and the recognized dialog data are treated as expert opinions.

5. The system of claim 4, wherein the dialog repair component processes user feedback to weight one expert opinion more than another expert opinion, and initiates confirmation processing based in part on the discrepancy data.

6. The system of claim 1, wherein the discrepancy detection component is integral to the user model prediction component.

7. The system of claim 6, wherein the dialog repair component looks for user feedback to update modifiable parameters and structures of the user model prediction component in an online fashion.

8. A computer-implemented method for leveraging discrepancies between user model predictions and speech recognition results for repairing dialog data, comprising:

processing speech data into predicted dialog data and recognized dialog data;

comparing the predictive dialog data and the recognized dialog data to generate difference data;

processing the difference data to determine a degree of difference between the predictive dialog data and the recognized dialog data; and

changing at least one of the predicted dialog data, the recognized dialog data, and a potential action based in part on optimization of a specific system action.

9. The method of claim 8, further comprising processing the difference data based on a confidence score that at least one of the predictive dialog data is correct, the recognized dialog data is correct or both the predictive dialog data and the recognized dialog data are incorrect.

10. The method of claim 8, further comprising generating weighted dialog data for at least one of the dialog data predicted by a user model and the dialog data recognized by a speech recognizer and initiating user confirmation to change the weighted dialog data.

11. The method of claim 8, further comprising processing the repaired dialog data into human understandable speech signals.

12. The method of claim 8, further comprising automatically updating reliability of at least one of the predicted dialog data and the recognized dialog data through user feedback as obtained for instance through a manual confirmation process.

13. The method of claim 8, further comprising deriving the predicted dialog data as a function of the received speech data.

14. The method of claim 8, further comprising updating a predictive user model based on a detected user response.

15. The method of claim 8, further comprising deriving the recognized dialog data based solely on acoustic information in the received speech data.

16. The method of claim 8, further comprising incorporating the difference data into an influence diagram to determine a degree of difference between the predictive dialog data and the recognized dialog data.

17. The method of claim 8, further comprising updating the reliability of at least one of a predictive user model and a speech recognizer according to feedback elicited from a user.

18. The method of claim 8, further comprising utilizing feedback to learn if a predictive user model or a speech recognizer is correct, wherein feedback occurs as an expected consequence of taking a system action and includes at least one of an explicit user action and a tacit acceptance of an action.

19. A computer-implemented system for repairing dialog data in a speech application, comprising:

computer-implemented means for identifying discrepancy data between predictive dialog data output from a predictive user model component and recognized dialog data output from a speech recognition component, the user model component and the speech recognition component each including a plurality of modifiable parameters and structures;

computer-implemented means for initiating a dialog repair process based in part on the discrepancy data; and

computer-implemented means for learning reliability of at least one of the user model prediction component and the speech recognition component based on user interaction data received as an expected consequence of taking a system action.

20. The system of claim 19, further comprising computer-implemented means for treating the dialog data output predicted by the predictive user model component and the dialog data output recognized by the speech recognition component as expert opinions, and computer-implemented means for utilizing user feedback elicited from the dialog repair process to weight one expert opinion more than another expert opinion.