SENTIMENT CLASSIFICATION USING OUT OF DOMAIN DATA

• US 20140101081A1
• Filed: 12/06/2013
• Published: 04/10/2014
• Est. Priority Date: 12/16/2008
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Ni, Xiaochuan, Chen, Zheng, Sun, Jian-Tao, Wang, Gang
• Granted Patent: US 8,942,470 B2
• Time in Patent Office: Days
• US Class Current: 706/12
• CPC Class Codes: G06F 16/353 into predefined classes
• CPC Class Codes: G06F 18/24 Classification techniques
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising
• CPC Class Codes: G06Q 30/0201 Market modelling; Market analysis; Collecting market data

Abstract

Providing sentiment classification of out of domain data are disclosed herein. In some aspects, a source domain having a trained classifier is matched to a target domain having a target classifier. The trained classifier may include identifiers that may be used to predict the sentiment of opinion data for the source domain. The target classifier may use the identifiers of the trained classifier to determine the sentiment of opinion data for the target domain.

Claims

1. A method of training a target classifier to categorize textual data, the method comprising:

matching a trained classifier to the target classifier, the trained classifier sharing at least one common attribute with the target classifier;

selecting identifiers from the trained classifier, the identifiers being predictors of a sentiment of the textual data as one of a positive opinion or a negative opinion; and

associating the identifiers with the target classifier.

2. The method of claim 1, wherein the textual data is opinion data that is posted on the Internet.

3. The method of claim 1, wherein the selecting the identifiers includes performing an analysis of terms in the textual data in view of a known sentiment of the textual data, wherein terms having a relatively high correlation with the sentiment are selected as identifiers.

4. The method of claim 3, wherein the known sentiment is available for a limited amount of opinion data.

5. The method of claim 3, wherein the identifiers are common identifiers that have a relatively high correlation with the sentiment of the textual data across multiple domains.

6. The method of claim 1, further comprising:

matching an additional trained classifier to the target classifier; and

selecting additional identifiers from the additional trained classifier to combine with the identifiers.

7. One or more computer-readable media storing computer-executable instructions that, when executed on one or more processors, performs acts comprising:

selecting a source domain that includes a trained classifier, the source domain being used to assist in training a target classifier of a target domain;

selecting identifiers from the source domain, the identifiers including domain identifiers, which have sentiment predictive capability of opinion data that is limited to the domain, and common identifiers, which have sentiment predictive capability of opinion data across multiple domains;

importing the common identifiers to the target domain; and

categorizing opinion data with the target classifier of the target domain using the common identifiers.

8. The one or more computer-readable media as recited in claim 7, wherein the acts further comprise outputting the categorized data as a sentiment summary of the opinion data.

9. The one or more computer-readable media as recited in claim 7, wherein the common identifiers are determined by comparing the identifiers of a first domain to at least one other domain, the common identifiers being identifiers that include the sentiment categorization between the domains.

10. The one or more computer-readable media as recited in claim 7, wherein the acts further comprise:

evaluating the common identifiers by categorizing opinion data with a known sentiment; and

removing common identifiers that do not accurately predict the known sentiment.

11. The one or more computer-readable media as recited in claim 7, wherein the acts further comprise selecting common identifiers from at least one other source domain to combine with the common identifiers from the source domain.

12. The one or more computer-readable media as recited in claim 11, wherein the identifiers are selected as portions of opinion data that have an accurate predictive ability to classify the opinion data by sentiment.

13. The one or more computer-readable media as recited in claim 7, wherein the opinion data is human-generated and made for consumption via the Internet.

14. A method, comprising:

selecting a source domain to train a target classifier, the source domain having identifiers that are used to predict the sentiment of opinion data of the source domain;

populating a target classifier of a target domain with the identifiers from the source domain; and

classifying opinion data of the target domain data using the target classifier by employing the identifiers to determine the sentiment of opinion data of the target domain.

15. The method as recited in claim 14, further comprising:

obtaining opinion entries of the source domain, the opinion data being labeled to include a known sentiment;

analyzing a portion of each of the opinion entries of the source domain with respect to the known sentiment; and

designating the analyzed portion as an identifier when the analyzed portion is determined to be an accurate predictor of the known sentiment.

16. The method as recited in claim 14, further comprising:

sorting opinion data of the target domain into groups with a designated sentiment based on the occurrence of the identifier in the opinion data;

locating additional identifiers in the sorted opinion data; and

assigning the sentiment of the identifier, which is used to create the sorted opinion data, to the additional identifier.

17. The method as recited in claim 16, further comprising assigning a first weight to the identifiers and a second weight to the additional identifiers.

18. The method as recited in claim 17, wherein the first weight is greater than the second weight.

19. The method as recited in claim 14, wherein the opinion data includes at least one of a customer review, a blog entry, or a forum entry.

20. The method as recited in claim 14, wherein selecting the source domain to train the target classifier includes selecting the source domain from multiple domains, where the selected source domain includes a strong correlation with the target domain, the correlation being measured at least in part by attributes of each domains.