Method and apparatus for auditing training supersets

• US 7,107,266 B1
• Filed: 10/24/2001
• Issued: 09/12/2006
• Est. Priority Date: 11/09/2000
• Status: Expired due to Term

• Current Assignee: SAP America Incorporated (SAP SE)
• Original Assignee: Inxight Software, Inc. (SAP SE)
• Inventors: Davis, Mark W., Breyman, Clark Leonard
• Primary Examiner(s): Le, Uyen
• Assistant Examiner(s): SHECHTMAN, CHERYL MARIA
• Application Number:
  US10/044,711
• Time in Patent Office: 1,784 Days
• Field of Search: 704/245, 706/12, 706/20, 706/21, 706/2, 707/5, 707/6
• US Class Current: 1/1
• CPC Class Codes: G06F 16/353 into predefined classes
• CPC Class Codes: Y10S 707/99935 Query augmenting and refining, e.g. inexact access
• CPC Class Codes: Y10S 707/99936 Pattern matching access

Abstract

The present invention includes methods and systems to audit and identify potential errors and/or omissions in a training set. Particular aspects of the present invention are described in the claims, specification and drawings.

Claims

1. A computer assisted method of auditing a superset of training data, the superset comprising examples of documents having one or more preexisting category assignments, the method including:

partitioning the superset into at least two disjoint sets, including a test set and a training set, wherein the test set includes one or more test documents and the training set includes examples of documents belonging to at least two categories;

automatically categorizing the test documents using the training set;

calculating a metric of confidence based on results of the categorizing step and comparing the automatic category assignments for the test documents to the preexisting category assignments; and

reporting the test documents and preexisting category assignments that are suspicious and the automatic category assignments that appear to be missing from the test documents, based on the metric of confidence.

2. The method of claim 1, further including repeating the partitioning, categorizing and calculating steps until at least one-half of the documents in the superset have been assigned to the test set.

3. The method of claim 2, wherein the test set created in the partition step has a single test document.

4. The method of claim 2, wherein the test set created in the partition step has a plurality of test documents.

5. The method of claim 1, further including repeating the partitioning, categorizing and calculating steps until more than 50 percent of the documents in the superset have been assigned to the test set.

6. The method of claim 5, wherein the partitioning, categorizing and calculating steps are carried out automatically.

7. The method of claim 5, wherein the partitioning, categorizing, calculating and reporting steps are carried out automatically.

8. The method of claim 1, wherein the partitioning, categorizing and calculating steps are carried out automatically.

9. The method of claim 1, wherein the partitioning, categorizing, calculating and reporting steps are carried out automatically.

10. The method of claim 1, wherein the categorizing step includes determining k nearest neighbors of the test documents and the calculating step is based on a k nearest neighbors categorization logic.

11. The method of claim 10, wherein the metric of confidence is an unweighted measure of distance between a particular test document and the examples of documents belonging to various categories.

12. The method of claim 11, where the unweighted measure includes application of a relationship

13. The method of claim 10, wherein the metric of confidence is a weighted measure of distance between a particular test document and the examples of documents belonging to various categories, the weighted measure taking into account the density of a neighborhood of the test document.

14. The method of claim 13, wherein the weighted measure includes application of a relationship

15. The method of claim 1, wherein the reporting step further includes filtering the reported test documents based on the metric of confidence.

16. The method of claim 15, wherein filtering further includes color coding the reported test documents based on the metric of confidence.

17. The method of claim 15, wherein filtering further includes selecting for display the reported test documents based on the metric of confidence.

18. The method of claim 1, wherein reporting includes generating a printed report.

19. The method of claim 1, wherein reporting includes generating a file conforming to XML syntax.

20. The method of claim 1, wherein reporting includes generating a sorted display identifying at least a portion of the test documents.

21. The method of claim 1, further including calculating a precision score for the reported test documents.

22. A computer assisted method of auditing a superset of training data, the superset comprising examples of documents having one or more preexisting category assignments, the method including:

determining k nearest neighbors of the documents in a test subset automatically partitioned from the superset;

automatically categorizing the documents based on the k nearest neighbors into a plurality of categories;

calculating a metric of confidence based on results of the categorizing step and comparing the automatic category assignments for the documents to the preexisting category assignments; and

reporting the documents in the test subset and preexisting category assignments that are suspicious and the automatic category assignments that appear to be missing from the documents in the test subset, based on the metric of confidence.

23. The method of claim 22, wherein the metric of confidence is an unweighted measure of distance between a particular test document and the examples of documents belonging to various categories.

24. The method of claim 23, where the unweighted measure includes application of a relationship

25. The method of claim 22, wherein the metric of confidence is a weighted measure of distance between a particular test document and the examples of documents belonging to various categories, the weighted measure taking into account the density of a neighborhood of the test document.

26. The method of claim 25, wherein the weighted measure includes application of a relationship

27. The method of claim 22, wherein the determining, categorizing and calculating steps are carried out automatically.

28. The method of claim 22, wherein the reporting step further includes filtering the documents based on the metric of confidence.

29. The method of claim 28, wherein the filtering step further includes color coding the reported documents based on the metric of confidence.

30. The method of claim 28, wherein the filtering step further includes selecting for display the reported documents based on the metric of confidence.

31. The method of claim 22, wherein reporting includes generating a printed report.

32. The method of claim 22, wherein reporting includes generating a file conforming to XML syntax.

33. The method of claim 22, wherein reporting includes generating a sorted display identifying at least a portion of the documents.

34. The method of claim 22, further including calculating a precision score for the reported documents.