INCORPORATING LEXICON KNOWLEDGE INTO SVM LEARNING TO IMPROVE SENTIMENT CLASSIFICATION

• US 20120271788A1
• Filed: 04/21/2011
• Published: 10/25/2012
• Est. Priority Date: 04/21/2011
• Status: Active Grant

• Current Assignee: Xerox Corporation (Xerox Holdings Corp.)
• Original Assignee: Palo Alto Research Center, Inc. (Xerox Holdings Corp.)
• Inventors: Fang, Ji, Chen, Bi
• Granted Patent: US 8,352,405 B2
• Time in Patent Office: Days
• US Class Current: 706/50
• CPC Class Codes: G06F 40/242 Dictionaries
• CPC Class Codes: G06F 40/247 Thesauruses; Synonyms
• CPC Class Codes: G06F 40/30 Semantic analysis

Abstract

A sentiment classifier for sentiment classification of content. An aspect classifier is configured to classify content as being related to a particular aspect of information, the aspect classifier incorporating at least a portion of the domain specific sentiment lexicon. A polarity classifier is then configured to classify the content classified by the aspect classifier as having one of a positive sentiment of the particular aspect of information, a negative sentiment of the particular aspect of information or as having no sentiment as to the particular aspect of information. The polarity classifier also incorporating at least a portion of the domain specific sentiment lexicon.

Claims

1. A sentiment classifier for sentiment classification of content comprising:

an aspect classifier configured to classify content as being related to a particular aspect of information, the aspect classifier incorporating at least a portion of a domain specific sentiment lexicon; and

a polarity classifier configured to classify the content classified by the aspect classifier as having one of a positive sentiment of the particular aspect of information, a negative sentiment of the particular aspect of information or as having no sentiment of the particular aspect of information, the polarity classifier incorporating at least a portion of the domain specific sentiment lexicon.

2. The sentiment classifier of claim 1 wherein the aspect classifier further incorporates a general purpose sentiment lexicon.

3. The sentiment classifier of claim 1 wherein the polarity classifier further incorporates a general purpose sentiment lexicon.

4. The sentiment classifier of claim 1 being implemented as a Support Vector Machine.

5. The sentiment classifier of claim 1 wherein the content is text configured in the form of sentences having words and/or phrases.

6. The sentiment classifier of claim 5 further configured to classify the content at the sentence level as being positive, negative or having no sentiment as to the particular aspect of information.

7. The sentiment classifier of claim 6 further configured to predict a main topic associated with the sentiment at the sentence level.

8. The sentiment classifier of claim 1 wherein the domain specific sentiment lexicon is configured of:

(i) domain specific words and/or phrases from an annotated corpus which has been filtered, (ii) domain specific words and/or phrases obtained via a search of the world wide web using a predetermined linguistic pattern and filtering the returned search results, and (iii) domain specific words and/or phrases obtained via a dictionary expansion technique.

9. The sentiment classifier of claim 8 wherein the domain specific lexicon includes a sub-lexicon classifying the domain specific words and/or phrases as being associated with one of the aspects of information and another sub-lexicon associating a sentiment to the classified one aspect of information.

10. A method of performing sentiment classification of content comprising:

classifying content as being related to a particular aspect of information by an aspect classifier, wherein the aspect classifier incorporates at least a portion of a domain specific sentiment lexicon; and

classifying the content classified by the aspect classifier as having one of a positive sentiment of the particular aspect of information, a negative sentiment of the particular aspect of information or as having no sentiment of the particular aspect of information by use of a polarity classifier, wherein the polarity classifier incorporates at least a portion of the domain specific sentiment lexicon.

11. The method of claim 10 wherein the aspect classifier further incorporates a general purpose sentiment lexicon.

12. The method of claim 10 wherein the polarity classifier further incorporates a general purpose sentiment lexicon.

13. The method of claim 10 wherein the aspect classifier and the polarity classifier are assembled together to form a sentiment classifier implemented as a Support Vector Machine.

14. The method of claim 10 wherein the content is text configured in the form of sentences having words and phrases.

15. The method of claim 14 further including classifying the content at the sentence level as being positive, negative or having no sentiment of the particular aspect of information.

16. The method of claim 15 further including predicting a main topic associated with the sentiment at the sentence level.

17. The method of claim 10 wherein the domain specific sentiment lexicon is configured by (i) obtaining domain specific words and/or phrases by filtering an annotated corpus, (ii) obtaining domain specific words and/or phrases by searching the world wide web via the internet using a predetermined linguistic pattern and filtering returned search results, and (iii) performing a dictionary expansion operation on domain specific words and/or phrases obtained via (i) and (ii).

18. The method of claim 17, wherein the words and/or phrases which have been filtered from the annotated corpus have been filtered by:

identifying aspects of information;

extracting the words and/or phrases from the annotated corpus in sentences labeled as expressing the identified aspects of information;

forming initial lists of lexicons corresponding to the identified aspects of information from the extracted words and/or phrases;

checking the words and/or phrases from the initial lists of lexicons for the identified aspects of information against the other initial lists of lexicon for the other aspects of information;

filtering out any words and/or phrases from the initial lists of lexicons for the aspects of information which matches words and/or phrases from any of the other initial lists of lexicons for the other aspects of information; and

generating filtered lists of lexicons for the aspects of information representing domain specific lexicons.

19. The method of claim 17 wherein content obtained via a search of the world wide web or internet is obtained by:

generating linguistic patterns corresponding to selected aspects of information;

selecting one of the generated linguistic patterns;

sending via a search engine the selected linguistic pattern to the world wide web or internet to obtain search results based on the selected linguistic pattern;

extracting words and/or phrases from the results returned from the search;

filtering out noise words and/or phrases by one of (i) removing predetermined stop words and/or phrases or (ii) using each of candidate content (words) as a new seed content (word) and redoing a search and when the original seed content (word) is retuned keeping the content (word), otherwise filtering out the content (word); and

adding the words and/or phrases that remain to the domain specific lexicon.

20. The method of claim 17 wherein the dictionary expansion includes finding at least one of synonyms and antonyms to words and/or phrases in the list, and adding at least one of the synonyms and antonyms to the list.