System and method for applying ranking SVM in query relaxation

• US 8,458,165 B2
• Filed: 06/28/2007
• Issued: 06/04/2013
• Est. Priority Date: 06/28/2007
• Status: Active Grant

• Current Assignee: Oracle International Corporation (Oracle Corporation)
• Original Assignee: Oracle International Corporation (Oracle Corporation)
• Inventors: Liao, Ciya, Chang, Thomas
• Primary Examiner(s): DWIVEDI, MAHESH H
• Application Number:
  US11/770,027
• Publication Number:
  US 20090006360A1
• Time in Patent Office: 2,168 Days
• Field of Search: 707/5, 707/713, 707/728, 707/737
• US Class Current: 707/713
• CPC Class Codes: G06F 16/284 Relational databases

Abstract

An enterprise-wide query relaxative support vector machine ranking algorithm approach provides enhanced functionality for query execution in a heterogeneous enterprise environment. Improved query results are obtained by adjusting ranking functions using machine learning methods to automatically train ranking functions. The improved query results are obtained using a list of document-query pairs that are modeled as a binary classification training problem, combination function which requires ranking and learning functions to be implemented representing document attributes and metadata utilizing query relaxation techniques and adjusted ranking functions. Machine learning methods implement user feedback to automatically train ranking functions.

Claims

1. A method in a computer system for providing improved search results in response to a user query, the method comprising:

obtaining a set of ranked query item pairs, each query item pair corresponding to a query and a hit document;

for each hit document;

executing a series of query relaxation operations, each query relaxation operation extracting one feature vector of a plurality of feature vectors for said each hit document; and

calculating a relevance score using the one extracted feature vector and a learned ranking function that is unique for that query relaxation operation; and

generating a hit list for the user query to be displayed to the user that contains hits from the executed query relaxation operations;

modeling a ranked item as a pair comprised of a query and a hit document and only items that have the same query, wherein each item is represented by a feature vector of the plurality of feature vectors, which lists features and corresponding feature weights configured to provide a user to tune a ranking function based on coupling a query relaxation method with a ranking support vector machine (SVM) application, and wherein the user determines an optimal fit between an initial list of document query hits and a revised list produced from a trained learning system machine learning function and a feature includes any attribute of a document that is used to determine a relevance of a document related to a given query.

2. A method according to claim 1, wherein obtaining a set of ranked query item pairs includes obtaining query item pairs from across a heterogeneous enterprise network.

3. A method according to claim 1, further comprising:

displaying the hit list to the user; and

receiving feedback from the user regarding the hit list, the feedback being used to adjust a weight for at least one ranking factor of a learned ranking function in response thereto.

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

allowing the user to continue to provide feedback in order to further update the learned ranking function until a desired hit is displayed that is consistent with user search result preferences.

5. A method according to claim 1, further comprising:

crawling a corpus of documents to produce an index of ranked item pairs.

6. A method according to claim 1, wherein:

at least one of the set of features includes an occurrence of a query term in metadata for a hit document, occurrence of a query term in a document body, and a document static rank.

7. A method according to claim 6, wherein:

the query term is at least one of a title, keyword, description, or reference text.

8. A method according to claim 6, wherein:

the static rank is at least one of a page rank and a human generated document static rank.

9. A method according to claim 1, further comprising:

terminating the execution of the series of query relaxation operations when a result page is filled by the hit list.

10. A method according to claim 1, wherein:

each feature weight is a linear combination of training vectors.

11. A method according to claim 1, further comprising:

classifying each query item pair into one of at least two classes based on a level of a relevancy score term.

12. A non-transitory computer-readable storage medium for providing improved search results in response to a user query having sets of instructions stored thereon which, when executed by a computer, cause the computer to:

obtain a set of ranked query item pairs, each query item pair corresponding to a query and a hit document;

perform for each hit document;

executing a series of query relaxation operations, each query relaxation operation extracting one feature vector of a plurality of feature vectors for said each hit document; and

calculating a relevance score using the one extracted feature vector and a learned ranking function that is unique for that query relaxation operation; and

generate a hit list for the user query to be displayed to the user that contains hits from the executed query relaxation operations;

model a ranked item as a pair comprised of a query and a hit document and only items that have the same query, wherein each item is represented by a feature vector of the plurality of feature vectors, which lists features and corresponding feature weights configured to provide a user to tune a ranking function based on coupling a query relaxation method with a ranking support vector machine (SVM) application, and wherein the user determines an optimal fit between an initial list of document query hits and a revised list produced from a trained learning system machine learning function and a feature includes any attribute of a document that is used to determine a relevance of a document related to a given query.

13. The non-transitory computer-readable storage medium according to claim 12, wherein the sets of instructions when further executed by the computer, cause the computer to obtain a set of ranked query item pairs includes obtaining query item pairs from across a heterogeneous enterprise network.

14. A non-transitory computer-readable storage medium according to claim 12, further comprising:

program code for displaying the hit list to the user; and

program code for receiving feedback from the user regarding the hit list, the feedback being used to adjust a weight for at least one ranking factor of a learned ranking function in response thereto.

15. A non-transitory computer-readable storage medium according to claim 14, further comprising:

program code for allowing the user to continue to provide feedback in order to further update the learned ranking function until a desired hit is displayed that is consistent with user search result preferences.

16. A non-transitory computer-readable storage medium according to claim 12, further comprising:

program code for crawling a corpus of documents to produce an index of ranked item pairs.

17. A non-transitory computer-readable storage medium according to claim 12, further comprising:

program code for terminating the execution of the series of query relaxation operations when a result page is filled by the hit list.

18. A system for providing improved search results in response to a user query, the system comprising:

a storage device; and

a computer processor in communication with the storage device, wherein the storage device has sets of instructions stored thereon which, when executed by the processor, cause the processor to;

obtain a set of ranked query item pairs, each query item pair corresponding to a query and a hit document;

for each hit document;

execute a series of query relaxation operations, each query relaxation operation extracting one feature vector of a plurality of feature vectors for said each hit document,; and

calculating a relevance score using the one extracted feature vector and a learned ranking function that is unique for that query relaxation operation; and

generate a hit list for the user query to be displayed to the user that contains hits from the executed query relaxation operations model a ranked item as a pair comprised of a query and a hit document and only items that have the same query, wherein each item is represented by a feature vector of the plurality of feature vectors, which lists features and corresponding feature weights configured to provide a user to tune a ranking function based on coupling a query relaxation method with a ranking support vector machine (SVM) application, and wherein the user determines an optimal fit between an initial list of document query hits and a revised list produced from a trained learning system machine learning function and a feature includes any attribute of a document that is used to determine a relevance of a document related to a given query.

19. A system according to claim 18, wherein the instructions, when executed by the processor, further cause the processor to:

display the hit list to the user; and

receive feedback from the user regarding the hit list, the feedback being used to adjust a weight for at least one ranking factor of a learned ranking function in response thereto.

20. A system according to claim 18, wherein the instructions, when executed by the processor, further cause the processor to:

terminate the execution of the series of query relaxation operations when a result page is filled by the hit list.