Apparatus and method for term context modeling for information retrieval

• US 20080052273A1
• Filed: 08/22/2006
• Published: 02/28/2008
• Est. Priority Date: 08/22/2006
• Status: Active Grant

• Current Assignee: Fujifilm Business Innovation Corp. (Fujifilm Holdings Corporation)
• Original Assignee: Fuji Xerox Company Limited (Xerox Holdings Corp.)
• Inventors: Pickens, Jeremy
• Granted Patent: US 7,937,397 B2
• Time in Patent Office: Days
• US Class Current: 707/3
• CPC Class Codes: G06F 16/3344 using natural language analysis
• CPC Class Codes: Y10S 707/917 Text

Abstract

A novel method for going beyond the observed properties of a keyword, to a model in which the presence of a term in a document is assessed not by looking at the actual occurrence of that term, but by a set of non-independent supporting terms, defining the context. In other words, similarity is determined not by properties of the keyword, but by properties of the keyword'"'"'s context. This yields a scoring for documents which is useful for ad hoc retrieval and, by extension, any information retrieval task where keyword-based similarity is needed.

Claims

1. A method for computing similarity between a first text object and a second text object, the method comprising:

a. Using the first text object to derive a context model associated with the first text object; and

b. Using the derived context model to compute similarity between the first text object and the second text object.

2. The method of claim 1, wherein the first text object comprises at least one keyword.

3. The method of claim 1, wherein the first text object comprises at least one phrase.

4. The method of claim 1, wherein the context model comprises at least one term predicting the occurrence of the first text object within the second text object.

5. The method of claim 1, wherein the context model comprises multiple terms predicting the occurrence of the first text object within the second text object.

6. The method of claim 1, wherein the context model comprises at least one phrase predicting the occurrence of the first text object within the second text object.

7. The method of claim 1, wherein the context model comprises multiple phrases predicting the occurrence of the first text object within the second text object.

8. The method of claim 1, wherein the context model comprises at least one keyword and at least one phrase predicting the occurrence of the first text object within the second text object.

9. The method of claim 1, wherein the context model comprises at least one term proximity condition predicting the occurrence of the first text object within the second text object.

10. The method of claim 1, wherein the context model comprises at least one term relationship condition predicting the occurrence of the first text object within the second text object.

11. The method of claim 1, wherein the similarity between the first text object and the second text object is computed using a probability score calculation.

12. The method of claim 1, further comprising retrieving at least one document based on the computed similarity.

13. The method of claim 12, further comprising ranking the retrieved documents based on the computed similarity.

14. The method of claim 1, wherein the first text object comprises a model term and the context model comprises a support feature helping to predict the model term.

15. The method of claim 14, wherein the context model further comprises a weight assigned to the support feature.

16. The method of claim 1, further comprising training the context model using a body of text.

17. A method for automatic induction of a context model for a term, the method comprising:

a. Selecting a feature set to contain only a feature for the term with no context;

b. Setting the initial weight to one for the feature;

c. Updating the weight for the feature; and

d. Performing feature induction.

18. The method of claim 17, wherein updating the weight for the feature comprises:

i. Calculating a gradient and adjusting the initial weight a small step in a direction of the gradient; and

ii. Repeating (a) until a predetermined condition is met.

19. The method of claim 17, wherein performing the feature induction comprises:

i. Enumerating a set of candidate features;

ii. For each of the set of the candidate features, computing an optimal weight;

iii. For each of the set of the candidate features, computing an information gain by adding each of the set of the candidate features to the feature set;

iv. Picking a candidate feature of the set of the candidate features, which yields a highest gain, and adding the candidate feature to the feature set; and

v. Returning to step (c), until a predetermined condition is met.

20. A computer programming product embodied on a computer readable medium, for computing similarity between a first text object and a second text object, the computer programming product comprising:

a. Code for using the first text object to derive a context model associated with the first text object; and

b. Code for using the derived context model to compute similarity between the first text object and the second text object.

21. The computer programming product of claim 20, wherein the first text object comprises at least one keyword.

22. The computer programming product of claim 20, wherein the first text object comprises at least one phrase.

23. The computer programming product of claim 20, wherein the context model comprises at least one term predicting the occurrence of the first text object within the second text object.

24. The computer programming product of claim 20, wherein the context model comprises multiple terms predicting the occurrence of the first text object within the second text object.

25. The computer programming product of claim 20, wherein the context model comprises at least one phrase predicting the occurrence of the first text object within the second text object.

26. The computer programming product of claim 20, wherein the context model comprises multiple phrases predicting the occurrence of the first text object within the second text object.

27. The computer programming product of claim 20, wherein the context model comprises at least one keyword and at least one phrase predicting the occurrence of the first text object within the second text object.

28. The computer programming product of claim 20, wherein the context model comprises at least one term proximity condition predicting the occurrence of the first text object within the second text object.

29. The computer programming product of claim 20, wherein the context model comprises at least one term relationship condition predicting the occurrence of the first text object within the second text object.

30. The computer programming product of claim 20, wherein the similarity between the first text object and the second text object is computed using a probability score calculation.

31. The computer programming product of claim 20, further comprising retrieving at least one document based on the computed similarity.

32. The computer programming product of claim 20, further comprising ranking the retrieved documents based on the computed similarity.

33. The computer programming product of claim 20, wherein the first text object comprises a model term and the context model comprises a support feature helping to predict the model term.

34. The computer programming product of claim 33, wherein the context model further comprises a weight assigned to the support feature.

35. The computer programming product of claim 20, further comprising training the context model using a body of text.

36. A computer programming product embodied on a computer readable medium, for automatic induction of a context model for a term, the computer programming product comprising:

a. Code for selecting a feature set to contain only a feature for the term with no context;

b. Code for setting the initial weight to one for the feature;

c. Code for updating the weight for the feature; and

d. Code for performing feature induction.

37. The computer programming product of claim 36, wherein code for updating the weight for the feature comprises:

i. Code for calculating a gradient and adjusting the initial weight a small step in a direction of the gradient; and

ii. Code for repeating (a) until a predetermined condition is met.

38. The computer programming product of claim 36, wherein code for performing the feature induction comprises:

i. Code for enumerating a set of candidate features;

ii. Code for computing an optimal weight for each of the set of the candidate features;

iii. Code for computing an information gain for each of the set of the candidate features, by adding each of the set of the candidate features to the feature set;

iv. Code for picking a candidate feature of the set of the candidate features, which yields a highest gain, and adding the candidate feature to the feature set;

v. Code for returning to step (c), until a predetermined condition is met.