Method and apparatus for training a translation disambiguation classifier

• US 20040254782A1
• Filed: 06/12/2003
• Published: 12/16/2004
• Est. Priority Date: 06/12/2003
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Li, Hang
• Granted Patent: US 7,318,022 B2
• Time in Patent Office: Days
• US Class Current: 704/8
• CPC Class Codes: G06F 40/30 Semantic analysis

Abstract

A method of training a classifier includes applying a first classifier to a first set of unlabeled data to form a first set of labeled data. The first classifier is able to assign data to classes in a first set of classes. A second classifier is applied to a second set of unlabeled data to from a second set of labeled data. The second classifier is able to assign data to classes in a second set of classes that is different from the first set of classes. The first and second sets of labeled data are used to retrain the first classifier.

Claims

1. A method of training a classifier, the method comprising:

applying a first classifier to a first set of unlabeled data to form a first set of labeled data, the first classifier capable of assigning data to classes in a first set of classes;

applying a second classifier to a second set of unlabeled data to form a second set of labeled data, the second classifier capable of assigning data to classes in a second set of classes that is different from the first set of classes; and

using the first set of labeled data and the second set of labeled data to retrain the first classifier to form a retrained classifier.

2. The method of claim 1 wherein the first set of unlabeled data comprises text in a first language and the second set of unlabeled data comprises text in a second language.

3. The method of claim 2 wherein the first set of classes comprises senses for a word in the first language.

4. The method of claim 3 wherein the second set of classes comprises senses for a word in the second language.

5. The method of claim 1 wherein using the first set of labeled data and the second set of labeled data to retrain the first classifier comprises:

determining a first probability component from the first set of labeled data;

determining a second probability component from the second set of labeled data; and

combining the first probability component and the second probability component to form a probability term used to define the first classifier.

6. The method of claim 5 wherein determining the first probability component comprises executing a first algorithm using the first set of labeled data and wherein determining the second probability component comprises executing a second algorithm that operates differently from the first algorithm using the second set of labeled data.

7. The method of claim 6 wherein the first algorithm comprises a maximum likelihood estimation algorithm.

8. The method of claim 7 wherein the second algorithm comprises an expectation maximization algorithm.

9. The method of claim 5 wherein combining the first probability component and the second probability component comprises linearly combining the first probability component and the second probability component.

10. The method of claim 1 further comprising using the first set of labeled data and the second set of labeled data to retrain the second classifier to form a second retrained classifier.

11. The method of claim 10 wherein using the first set of labeled data to retrain the first classifier comprises executing a first algorithm on the first set of labeled data and wherein using the first set of labeled data to retrain the second classifier comprises executing a second algorithm on the first set of labeled data, where the second algorithm operates differently than the first algorithm.

12. The method of claim 11 wherein the first algorithm comprises a maximum likelihood estimation algorithm.

13. The method of claim 12 wherein the second algorithm comprises an expectation maximization algorithm.

14. The method of claim 10 further comprising:

applying the retrained classifier to a set of unlabeled data to form a first additional set of labeled data;

applying the second retrained classifier to a set of unlabeled data to form a second additional set of labeled data; and

using the first additional set of labeled data and the second additional set of labeled data to retrain the retrained classifier.

15. A computer-readable medium having computer-executable instructions for performing steps comprising:

generating first language labeled data that indicates a sense of at least one word in a first language;

generating second language labeled data that indicates a sense of at least one word in a second language; and

using the first language labeled data and the second language labeled data to train a classifier for the first language, where the classifier can be used to identify a sense of a word in the first language.

16. The computer-readable medium of claim 15 wherein generating first language labeled data comprises applying unlabeled data in the first language to a classifier for the first language.

17. The computer-readable medium of claim 16 wherein training a classifier for the first language comprises retraining the classifier used to generate the first language labeled data.

18. The computer-readable medium of claim 16 wherein generating second language labeled data comprises applying unlabeled data in the second language to a classifier for the second language.

19. The computer-readable medium of claim 15 wherein using the first language labeled data and the second language labeled data to train a classifier comprises:

determining a first probability component using the first language labeled data;

determining a second probability component using the second language labeled data; and

combining the first probability component and the second probability component to form a probability term for the classifier.

20. The computer-readable medium of claim 19 wherein combining the first probability component and the second probability component comprises linearly combining the first probability component and the second probability component.

21. The computer-readable medium of claim 19 wherein determining the first probability component comprises using a first algorithm and wherein determining the second probability component comprises using a second algorithm that is different from the first algorithm.

22. The computer-readable medium of claim 21 wherein the first algorithm is a maximum likelihood estimation algorithm.

23. The computer-readable medium of claim 22 wherein the second algorithm is an expectation maximization algorithm.

24. The computer-readable medium of claim 15 further comprising using the classifier trained from the first language labeled data and the second language labeled data to classify unlabeled data in the first language to form additional first language labeled data.

25. The computer-readable medium of claim 15 further comprising using the first language labeled data and the second language labeled data to train a classifier for the second language.

26. The computer readable medium of claim 25 wherein using the first language labeled data and the second language labeled data to train a classifier for the second language comprises:

determining a first probability component using the first language labeled data;

determining a second probability component using the second language labeled data; and

combining the first probability component and the second probability component to form a probability term for the classifier for the second language.

27. The computer-readable medium of claim 26 wherein combining the first probability component and the second probability component comprises linearly combining the first probability component and the second probability component.

28. The computer-readable medium of claim 26 wherein determining the first probability component comprises using a first algorithm and wherein determining the second probability component comprises using a second algorithm that is different from the first algorithm.

29. The computer-readable medium of claim 28 wherein the second algorithm is a maximum likelihood estimation algorithm.

30. The computer-readable medium of claim 29 wherein the first algorithm is- an expectation maximization algorithm.

31. The computer-readable medium of claim 25 further comprising using the classifier for the second language to classify unlabeled data in the second language to form additional second language labeled data.