Apparatus and method for model adaptation for spoken language understanding

• US 7,996,219 B2
• Filed: 03/21/2005
• Issued: 08/09/2011
• Est. Priority Date: 03/21/2005
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: AT&T Intellectual Property I LP (AT&T, Inc.)
• Inventors: Tur, Gokhan
• Primary Examiner(s): Opsasnick, Michael N
• Application Number:
  US11/085,587
• Publication Number:
  US 20060212293A1
• Time in Patent Office: 2,332 Days
• Field of Search: 704/244
• US Class Current: 704/244
• CPC Class Codes: G10L 15/065 Adaptation

Abstract

An apparatus and a method are provided for building a spoken language understanding model. Labeled data may be obtained for a target application. A new classification model may be formed for use with the target application by using the labeled data for adaptation of an existing classification model. In some implementations, the existing classification model may be used to determine the most informative examples to label.

Claims

1. A method of building a classification model for a target application, comprising:

obtaining existing labeled data from an existing classification model of a first domain-specific application;

obtaining data labeled for the target application; and

forming, via a processor, a new classification model for use with the target application by using the existing labeled data from the existing classification model and the data labeled for the target application to train the new classification model, the target application being different from the first domain-specific application and in a different domain from the first domain-specific application.

2. The method of claim 1, wherein the act of forming a new classification model for use with the target application by using the labeled data for adaptation of the existing classification model further comprises:

determining a distance from the existing classification model to the new classification model; and

performing the adaptation of the existing model by using the distance.

3. The method of claim 2, wherein the classification model is a speech processing model and wherein the distance comprises a logistic loss function.

4. The method of claim 1, further comprising:

predicting labels of data for the target application by using the existing classification model;

determining which ones of the data are most informative for labeling; and

presenting, to a labeler, the determined most informative ones of the data for labeling.

5. The method of claim 1, further comprising:

predicting labels of a plurality of utterances for the target application by using the existing classification model;

determining which ones of the plurality of utterances are most informative for labeling; and

presenting, to a labeler, the determined ones of the plurality of utterances for labeling.

6. The method of claim 5, wherein the act of determining which ones of the plurality of utterances are most informative for labeling further comprises:

determining a confidence score with respect to each of the predicted labels of the plurality of utterances by using the existing classification model; and

determining the ones of the plurality of utterances that are most informative for labeling by determining ones of the plurality of utterances that have the lowest determined confidence scores.

7. An apparatus comprising:

a processor; and

storage for storing instructions for the processor, wherein the apparatus is configured to;

obtain existing labeled data from an existing classification model of a first domain-specific application;

obtain data labeled for a target application; and

form a new classification model for use with the target application by using the existing labeled data from the existing classification model and the data labeled for the target application to train the new classification model, the target application being different from the first domain-specific application and in a different domain from the first domain-specific application.

8. The apparatus of claim 7, wherein when the apparatus forms a new classification model for use with the target application by using the labeled data for adaptation of an existing classification model, the apparatus is configured to:

determine a distance from the existing classification model to the new classification model; and

perform the adaptation of the existing model by using the distance.

9. The apparatus of claim 8, wherein the classification model is a speech processing model and wherein the distance comprises a logistic loss function.

10. The apparatus of claim 7, wherein the apparatus is further configured to:

predict labels of a plurality of utterances for the target application by using the existing classification model;

determine which ones of the plurality of utterances are most informative for labeling; and

present, to a labeler, the determined ones of the plurality of utterances for labeling.

11. The apparatus of claim 10, wherein when the apparatus determines which ones of the plurality of utterances are most informative for labeling, the apparatus is configured to:

determine a confidence score with respect to each of the predicted labels of the plurality of utterances by using the existing classification model; and

determine the ones of the plurality of utterances that are most informative for labeling by determining ones of the plurality of utterances that have the lowest determined confidence scores.

12. A non-transitory machine-readable medium having instructions, stored therein, for a processor, the machine-readable medium comprises:

instructions for inputting existing labeled data from an existing classification model of a first domain-specific application;

instructions for inputting data labeled for a target application; and

instructions for forming a new classification model for use with the target application by using the existing labeled data from the existing classification model and the data labeled for the target application to train the new classification model, the target application being different from the first domain-specific application and in a different domain from the first domain-specific application.

13. The non-transitory machine-readable medium of claim 12, wherein the instructions for forming a new classification model for use with the target application by using the labeled data for adaptation of an existing classification model comprises:

instructions for determining a distance from the existing classification model to the new classification model; and

instructions for performing the adaptation of the existing model by using the distance.

14. The non-transitory machine-readable medium of claim 13, wherein the classification model is a speech processing model and wherein the distance comprises a logistic loss function.

15. The non-transitory machine-readable medium of claim 12, further comprising:

instructions for predicting labels of a plurality of utterances for the target application by using the existing classification model;

instructions for determining which ones of the plurality of utterances are most informative for labeling; and

instructions for presenting, to a labeler, the determined ones of the plurality of utterances for labeling.

16. The non-transitory computer-readable medium of claim 15, wherein the instructions for determining which ones of the plurality of utterances are most informative for labeling further comprises:

instructions for determining a confidence score with respect to each of the predicted labels of the plurality of utterances by using the existing classification model; and

instructions for determining the ones of the plurality of utterances that are most informative for labeling by determining ones of the plurality of utterances that have the lowest determined confidence scores.

17. An apparatus comprising:

means for obtaining existing labeled data from an existing classification model for a first domain-specific application;

means for obtained data labeled for a target application; and

means for forming a new classification model for use with the target application by using the existing labeled data from the existing classification model and the data labeled for the target application to train the new classification model, the target application being different from the first domain-specific application and in a different domain from the first domain-specific application.

18. The apparatus of claim 17, wherein the means for forming a new classification model for use with the target application by using the labeled data for adaptation of an existing classification model comprises:

means for determining a distance from the existing classification model to the new classification model; and

means for performing the adaptation of the existing model by using the distance.

19. The apparatus of claim 17, further comprising:

means for predicting labels of a plurality of utterances for the target application by using the existing classification model;

means for determining which ones of the plurality of utterances are most informative for labeling; and

means for presenting, to a labeler, the determined ones of the plurality of utterances for labeling.

20. The apparatus of claim 19, wherein the means for determining which ones of the plurality of utterances are most informative for labeling further comprises:

means for determining a confidence score with respect to each of the predicted labels of the plurality of utterances by using the existing classification model; and

means for determining the ones of the plurality of utterances that are most informative for labeling by determining ones of the plurality of utterances that have the lowest determined confidence scores.