TRANSFER OF LEARNING FOR QUERY CLASSIFICATION

• US 20120259801A1
• Filed: 04/06/2011
• Published: 10/11/2012
• Est. Priority Date: 04/06/2011
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Ji, Lei, Yan, Jun, Liu, Ning, Chen, Zheng
• Granted Patent: US 8,719,192 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06N 20/00 Machine learning

Abstract

Transfer of learning trains a new domain for the classification of search queries according to different tasks, as well as the generation of a corresponding domain-specific query classifier that may be used to classify the search queries according to the different tasks in the new domain. The transfer of learning may include preparing a new domain to receive classification knowledge from one or more source domains by populating the new domain with preliminary query patterns extracted for a search engine log. The transfer of learning may further include preparing the classification knowledge in each source domain for transfer to the new domain. The classification knowledge in each source domain may then be transferred to the new domain.

Claims

1. A computer-readable medium storing computer-executable instructions that, when executed, cause one or more processors to perform acts comprising:

preparing a new domain to receive classification knowledge from one or more source domains by populating the new domain with preliminary query patterns extracted from a search engine log;

preparing the classification knowledge in each of the one or more source domains for transfer to the new domain; and

transferring the classification knowledge from the one or more source domains to the new domain.

2. The computer-readable medium of claim 1, further comprising instructions that, when executed, cause the one or more processors to perform acts comprising:

selecting features for classifier training from the classification knowledge transferred to the new domain; and

constructing a classifier based on the features selected for classifier training.

3. The computer-readable medium of claim 2, further comprising instructions that, when executed, cause the one or more processors to perform an act of classifying a search query into the new domain using the classifier.

4. The computer-readable medium of claim 1, wherein the preparing the classification knowledge in each of the one or more source domains include:

extracting seed query patterns for a particular source domain from the search engine log;

labeling each of the seed query patterns with an associated task classification and an associated task probability that represent a probability that a corresponding seed query pattern belongs to the associated task classification; and

generating additional query patterns for the particular source domain based at least on the seed query patterns to form an expanded set of query patterns, each of the additional query patterns having a corresponding task classification and a corresponding task probability.

5. The computer-readable medium of claim 4, wherein the preparing the classification knowledge in each of the one or more source domains includes refining the expanded set of query patterns by discarding at least one query pattern of the expanded set of query patterns with a task probability that is below a probability threshold.

6. The computer-readable medium of claim 4, wherein the transferring includes:

minimizing at least one objective function to select query patterns from the expanded set of query patterns of a source domain for the new domain based at least on consistency between the query patterns selected and the preliminary query patterns; and

transferring the selected query patterns to the new domain, each of the selected query patterns having an associated task classification and an associated task probability that represent a probability that a corresponding selected query pattern belongs to the associated task classification.

7. The computer-readable medium of claim 6, wherein the objective function is expressed as:

8. The computer-readable medium of claim 6, further comprising generating additional query patterns for the new domain based at least on the selected query patterns to form an enlarged set of query patterns.

9. The computer-readable medium of claim 8, further comprising refining the enlarged set of query patterns by discarding at least one query pattern of the enlarged set of query patterns with a task probability that is below a probability threshold.

10. The computer-readable medium of claim 1, wherein transferring includes:

computing a similarity distance between each of a plurality of source domains and the new domain;

discarding at least one of the plurality of source domains with a similarity distance that is farther than a predefined distance threshold;

minimizing a plurality of objective functions to select one or more query patterns of the plurality of source domains for the new domain based at least on consistency between the one or more query patterns and the preliminary query patterns; and

transferring the one or more selected query patterns to the new domain, each of the selected query patterns having an associated task classification and an associated task probability that represent a probability that a corresponding selected query pattern belongs to the associated task classification.

11. The computer-readable medium of claim 10, wherein the transferring further includes:

generating additional query patterns for the new domain based at least on the selected query patterns to form an enlarged set of query patterns, each of the additional query patterns having a corresponding task classification and a corresponding task probability; and

refining the enlarged set of query patterns by discarding at least one query pattern of the enlarged set of query patterns with a task probability that is below a probability threshold.

12. A computer implemented method, comprising:

under control of one or more computing devices configured with executable instructions,

extracting seed query patterns for a source domain from a search engine log;

labeling each of the seed query patterns with an associated task classification and an associated task probability that represent a probability that a corresponding seed query pattern belongs to the associated task classification; and

generating additional query patterns for the source domain based at least on the seed query patterns to form an expanded set of query patterns, each of the additional query patterns having a corresponding task classification and a corresponding task probability; and

transferring one or more resultant query patterns from the expanded set of query patterns to a new domain.

13. The computer implemented method of claim 12, further comprising refining the expanded set of query patterns by discarding at least one query pattern of the expanded set of query patterns with a task probability that is below a probability threshold.

14. The computer implemented method of claim 12, further comprising preparing the new domain to receive a plurality of query patterns from the expanded set of query patterns from the source domain by populating the new domain with preliminary query patterns extracted for a search engine log.

15. The computer implemented method of claim 14, further comprising minimizing an objective function to select the plurality of query patterns from the expanded set of query patterns for the new domain based at least on consistency between the one or more resultant query patterns and the preliminary query patterns.

16. The computer implemented method of claim 15, further comprising:

generating additional query patterns for the new domain based at least on the plurality of query patterns using a Random Walk model algorithm to form an enlarged set of query patterns; and

refining the enlarged set of query patterns by discarding at least one query pattern of the enlarged set of query patterns with a task probability that is below a probability threshold.

17. The computer implemented method of claim 16, further comprising:

selecting features for classifier training from the enlarged set of query patterns; and

constructing a classifier based on the features selected for classifier training.

18. A system, comprising:

one or more processors; and

a memory that includes a plurality of computer-executable modules, the plurality of computer-executable components comprising;

a domain preparation module that prepares a new domain to receive refined query patterns associated with one or more source domains by populating the new domain with preliminary query patterns extracted from a search engine log;

a source domain module that produces the refined query patterns associated with each of the one or more source domains for transfer to the new domain; and

a learning transfer module that transfers one or more of the refined query patterns from the one or more source domains to the new domain.

19. The system of claim 18, wherein the source domain module produces the refined query patterns at least by:

extracting seed query patterns for a particular source domain from the search engine log;

labeling each of the seed query patterns with an associated task classification; and

generating additional query patterns for the particular source domain based at least on the seed query patterns to form the refined query patterns, each of the additional query patterns having a corresponding task classification.

20. The system of claim 18, wherein the learning transfer module transfers the refined query patterns at least by:

minimizing at least one objective function to select the one or more refined query patterns of a particular source domain for the new domain based at least on consistency between the one or more refined query patterns and the preliminary query patterns; and

transferring the one or more refined query patterns to the new domain, each of the one or more refined query patterns having an associated task classification and an associated task probability that represent a probability that a corresponding refined query pattern belongs to the associated task classification.