Knowledge re-use for call routing

• US 8,732,114 B2
• Filed: 04/13/2009
• Issued: 05/20/2014
• Est. Priority Date: 04/14/2008
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Nuance Communications, Inc. (Microsoft Corporation)
• Inventors: Liu, Ding
• Primary Examiner(s): Chen, Alan
• Assistant Examiner(s): BUSS, BENJAMIN J
• Application Number:
  US12/422,671
• Publication Number:
  US 20090259613A1
• Time in Patent Office: 1,863 Days
• Field of Search: 706/55, 704/9, 704/257, 704/275
• US Class Current: 706/55
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G10L 15/1815 Semantic context, e.g. disambiguation of the recognition hypotheses based on word meaning
• CPC Class Codes: G10L 15/22 Procedures used during a speech recognition process, e.g. man-machine dialogue

Abstract

A method is described for semantic classification in human-machine dialog applications, for example, call routing. Utterances in a new training corpus of a new semantic classification application are tagged using a pre-existing semantic classifier and associated pre-existing classification tags trained for an earlier semantic classification application.

Claims

1. A method for semantic classification comprising:

producing a labeled training corpus by labeling utterances in an unlabeled training corpus of a new semantic classification application using at least two pre-existing semantic classifiers that have been trained for different earlier semantic classification applications, each of the at least two pre-existing semantic classifiers having respective associated pre-existing classification tags that were trained for the different earlier semantic classification applications; and

training a new semantic classifier on the labeled training corpus to produce a set of new classification tags.

2. A method according to claim 1, training the new semantic classifier comprises:

training the new semantic classifier on the labeled training corpus to produce a set of new classification tags for the new semantic classifier including a plurality of the pre-existing classification tags.

3. A method according to claim 2, further comprising:

creating a set of application-specific company-generic classification tags from the set of new classification tags.

4. A method according to claim 1, wherein at least one of the pre-existing semantic classifiers or the new semantic classifier is a statistical classifier.

5. A method according to claim 1, wherein at least one of the earlier semantic classification applications or the new semantic classification application is a call routing application.

6. A method according to claim 1, wherein the new semantic classification application and the earlier semantic classification applications are in different application domains.

7. A method according to claim 1, wherein the new semantic classification application and one of the earlier semantic classification applications are in the same application domain.

8. A method of abstracting a semantic classifier, the method comprising:

producing a labeled training corpus by labeling utterances in an unlabeled training corpus of a new semantic classification application using at least two pre-existing semantic classifiers that have been trained for different earlier semantic classification applications, each of the at least two pre-existing semantic classifiers having respective associated pre-existing classification tags that were trained for the different earlier semantic classification applications;

training a first new semantic classifier on the labeled training corpus to produce a set of new classification tags;

identifying, by a computer system, application-specific features associated with an application-specific semantic classifier in the set of new classification tags;

replacing the application-specific features with generic stem rules in the set of new classification tags to create application-specific, company-generic classification tags for a new semantic classification application; and

replacing the application-specific semantic classifier with a generic semantic classifier based on the generic stem rules and the application-specific, company-genetic classification tags.

9. A method according to claim 8, further comprising:

using the application-specific semantic classifier and the generic stem rules to implement a new application-specific semantic classifier.

10. A method according to claim 9, wherein the semantic classifiers are in different application domains.

11. A method according to claim 9, wherein the semantic classifiers are in the same application domain.

12. A method according to claim 9, wherein the semantic classifiers are statistical classifiers.

13. A method according to claim 8, wherein at least one of the semantic classifiers is a call routing classifier.

14. A computer program product implemented in a non-transitory computer readable storage medium for semantic classification, the product comprising:

program code for producing a labeled training corpus by labeling utterances in an unlabeled training corpus of a new semantic classification application using at least two pre-existing semantic classifiers that have been trained for different earlier semantic classification applications, each of the at least two pre-existing semantic classifiers having respective associated pre-existing classification tags that were trained for the different earlier semantic classification applications; and

program code for training a new semantic classifier on the labeled training corpus to produce a set of new classification tags.

15. A product according to claim 14, wherein the program code for training the new semantic classifier comprises:

program code for training the new semantic classifier on the labeled training corpus to produce a set of new classification tags for the new semantic classifier including a plurality of the pre-existing classification tags.

16. A product according to claim 15, further comprising:

program code for creating a set of application-specific company-generic classification tags from the set of new classification tags.

17. A product according to claim 14, wherein at least one of the pre-existing semantic classifiers or the new semantic classifier is a statistical classifier.

18. A product according to claim 14, wherein at least one of the earlier semantic classification applications or the new semantic classification application is a call routing application.

19. A product according to claim 14, wherein the new semantic classification application and the earlier semantic classification applications are in different application domains.

20. A product according to claim 14, wherein the new semantic classification application and one of the earlier semantic classification applications are in the same application domain.

21. A system for developing a new semantic classification application based on earlier semantic classification applications, the system comprising:

means for labeling utterances in an initially unlabeled training corpus of the new semantic classification application using a first pre-existing semantic classifier trained for a first earlier semantic classification application;

means for labeling utterances in the training corpus of the new semantic classification application using first pre-existing classification tags that are associated with the first pre-existing semantic classifier and that have been trained for the first earlier semantic classification application;

means for labeling utterances in the training corpus of the new semantic classification application using a second pre-existing semantic classifier trained for a second earlier semantic classification application; and

means for labeling utterances in the training corpus of the new semantic classification application using second pre-existing classification tags that are associated with the second pre-existing semantic classifier and that have been trained for the second earlier semantic classification application; and

means for training a new semantic classifier on the training corpus to produce a set of new classification tags.

22. A system according to claim 21, further comprising:

means for training the new semantic classifier on the training corpus to produce a set of new classification tags for the new semantic classifier including a plurality of the pre-existing classification tags.

23. A system according to claim 22, further comprising:

means for creating a set of application-specific company-generic classification tags from the set of new classification tags.

24. A system according to claim 21, wherein at least one of the first pre-existing semantic classifier, the second pre-existing semantic classifier, or the new semantic classifier is a statistical classifier.

25. A system according to claim 21, wherein at least one of the first earlier semantic classification application, the second earlier semantic classification application, or the new semantic classification application is a call routing application.

26. A system according to claim 21, wherein the new semantic classification application and at least one of the first earlier semantic classification application or the second earlier semantic classification application are in different application domains.

27. A system according to claim 21, wherein the new semantic classification application and either the first earlier semantic classification application or the second earlier semantic classification application are in the same application domain.