Topical sentiments in electronically stored communications

• US 20060069589A1
• Filed: 09/30/2005
• Published: 03/30/2006
• Est. Priority Date: 09/30/2004
• Status: Active Grant

• Current Assignee: Buzzmetrics, Ltd. (Nielsen Holdings plc)
• Original Assignee: Buzzmetrics, Ltd. (Nielsen Holdings plc)
• Inventors: Hurst, Matthew F., Nigam, Kamal P.
• Granted Patent: US 7,523,085 B2
• Time in Patent Office: Days
• US Class Current: 706/55
• CPC Class Codes: G06F 40/253 Grammatical analysis; Style critique
• CPC Class Codes: G06Q 30/02 Marketing; Price estimation or determination; Fundraising

Abstract

The present application presents methods for performing topical sentiment analysis on electronically stored communications employing fusion of polarity and topicality. The present application also provides methods for utilizing shallow NLP techniques to determine the polarity of an expression. The present application also provides a method for tuning a domain-specific polarity lexicon for use in the polarity determination. The present application also provides methods for computing a numeric metric of the aggregate opinion about some topic expressed in a set of expressions.

Claims

1. A computer implemented method for obtaining topical sentiments from a plurality of electronically stored communications comprising the steps of:

determining with the assistance of a computer whether each communication in a plurality of communications is topical to a first predefined topic;

for each communications determined to be topical to the predefined topic, separating with the assistance of a computer the communication into one or more expressions;

for each expression, determining with the assistance of a computer if the expression is topical to a second predefined topic; and

for each expression that is determined to be topical to the second predefined topic, determining with the assistance of a computer a polarity of the expression.

2. The method of claim 1, wherein the polarity is taken from a group including at least the following sentiments:

positive, negative, and neutral.

3. The method of claims 1, wherein the step of determining the polarity of the expression includes the steps of:

establishing a topical domain being explored;

generating a polarity lexicon of sentimental words and/or phrases associated with the topical domain;

utilizing with the assistance of a computer the polarity lexicon against words and/or phrases found in the expression; and

assigning at least one polar phrase in the expression a polarity associated with a matching word and/or phrase in the polarity lexicon.

4. The method of claims 1, wherein the step of determining the polarity of the expression further comprises the step of analyzing with the assistance of a computer the expression with syntactic and/or semantic rules.

5. The method of claims 1, wherein the step of determining with the assistance of a computer whether each communication in a plurality of communications is topical to a first predefined topic includes the step of processing each communication with a text classifier.

6. The method of claim 5, wherein the text classifier utilizes an algorithm taken from a group consisting of:

a Winnow algorithm, a Support Vector Machine algorithm, a k-Nearest Neighbor algorithm and a rules-based classifier.

7. The method of claims 1, wherein the step of, for each expression, determining with the assistance of a computer if the expression is topical to a second predefined topic includes the step of processing each expression with a text classifier.

8. The method of claim 7, wherein the text classifier utilizes an algorithm taken from a group consisting of:

a Winnow algorithm, a Support Vector Machine algorithm, a k-Nearest Neighbor algorithm and a rules-based classifier.

9. The method of claim 1, further comprising the step of calculating with the assistance of a computer an aggregate metric from the plurality of expressions which estimates the frequency of positive and/or negative polar expressions.

10. The method of claim 9, wherein the step of calculating an aggregate metric from the plurality of expressions includes the generation of statistically-valid confidence bounds on the aggregate metric.

11. The method of claim 10, wherein the step of calculating an aggregate metric from the plurality of expressions includes the steps of:

for each of the plurality of expressions, estimating an opinion based upon the presence, absence or strength of polarity associated with the predefined topic; and

aggregating the overall opinion for the plurality of expressions.

12. The method of claim 11, wherein the step of aggregating the overall opinion for the plurality of expressions includes a step of normalizing the ratio of empirical or estimated frequency of positive and negative polarity associated with the predefined topic.

13. The method of claim 10 wherein the step of calculating an aggregate metric from the plurality of expressions includes utilizing Bayesian statistics to derive estimates for positive and negative frequencies of polar expressions.

14. The method of claim 1, wherein:

the first predefined topic is a general topic and the second predefined topic is a specific topic associated with the general topic.

15. The method of claim 14, wherein the general topic is a product or service and the specific topic is a feature of the product or service.

16. The method of claim 14, wherein the general topic is a commercial brand and the specific topic is a feature of the commercial brand.

17. The method of claim 1, wherein the first predefined topic and the second predefined topic are the same topic.

18. A computer implemented method for calculating, from a plurality of electronically stored expressions, an aggregate metric which estimates a frequency of positive and/or negative polar expressions contained in the expressions, comprising the steps of:

for each of a plurality of electronically stored expressions, determining with the assistance of a computer an opinion contained in the expressions based upon at least one of the presence, absence and strength of polarity associated with a predefined topic; and

calculating an aggregate metric from the determined opinions of the plurality of expressions.

19. The method of claim 18, wherein the step of calculating an aggregate metric from the determined opinions of the plurality of expressions includes the generation of statistically-valid confidence bounds on the aggregate metric.

20. The method of claim 18, wherein the step of calculating an aggregate metric from the determined opinions of the plurality of expressions includes a step of normalizing the ratio of empirical or estimated frequency of positive and negative polarity associated with the predefined topic.

21. The method of claim 20 wherein the step of calculating an aggregate metric from the determined opinions of the plurality of expressions further includes utilizing Bayesian statistics to derive estimates for positive and negative frequencies of polar expressions.

22. The method of claim 18, wherein at least a portion of the plurality of expressions are taken from larger electronically stored communications.

23. The method of claim 18, wherein the step of calculating an aggregate metric from the determined opinions of the plurality of expressions includes utilizing Bayesian statistics to derive estimates for positive and negative frequencies of polar expressions.

24. The method of claim 18, wherein the step of determining an opinion contained in the expressions includes the steps of, for each expression:

determining with the assistance of a computer that the expression is topical to the predefined topic; and

determining with the assistance of a computer a polarity of the expression.

25. The method of claim 24, wherein the step of determining with the assistance of a computer that the expression is topical to the predefined topic includes the step of utilizing a text classifier.

26. The method of claim 25, wherein the step of determining with the assistance of a computer a polarity of the expression includes the steps of:

establishing a topical domain being explored;

generating a polarity lexicon of sentimental words and/or phrases associated with the topical domain;

utilizing with the assistance of a computer the polarity lexicon against words and/or phrases found in the expression; and

assigning at least one polar phrase in the expression a polarity associated with a matching word and/or phrase in the polarity lexicon.

27. The method of claim 25, wherein the step of determining with the assistance of a computer a polarity of the expression includes the step of analyzing with the assistance of a computer the expression with syntactic and/or semantic rules.

28. The method of claim 24, wherein the step of determining with the assistance of a computer a polarity of the expression includes the steps of:

establishing a topical domain being explored;

generating a polarity lexicon of sentimental words and/or phrases associated with the topical domain;

utilizing with the assistance of a computer the polarity lexicon against words and/or phrases found in the expression; and

assigning at least one polar phrase in the expression a polarity associated with a matching word and/or phrase in the polarity lexicon.

29. The method of claim 24, wherein the step of determining with the assistance of a computer a polarity of the expression includes the step of analyzing with the assistance of a computer the expression with syntactic and/or semantic rules.

30. A computer implemented method for finding one or more polar expressions in an electronically stored communication, comprising the step of analyzing with the assistance of a computer the electronically stored communication for one or more polar expressions within the electronically stored communication, wherein the analyzing step includes the steps of:

providing a polarity lexicon of sentimental words and/or phrases associated with a topical domain;

utilizing with the assistance of a computer the polarity lexicon against words and/or phrases found in the expression; and

assigning with the assistance of a computer at least one word/phrase in the expression a polarity associated with a matching word/phrase in the polarity lexicon.

31. The method of claim 30, wherein the step of assigning with the assistance of a computer at least one word/phrase in the expression a polarity associated with a matching word/phrase in the polarity lexicon, includes the steps of:

separating the expression into word/phrase chunks;

tagging the separated word/phrase chunks with part-of-speech tags;

applying the polarity lexicon against the word/phrase chunks to tag one or more of the word/phrase chunks with a polarity tag;

applying syntactic and semantic rules against the tagged word/phrase chunks to elevate the polarity of the word/phrase chunk to the entire expression.

32. The method of claim 31, wherein the step of applying syntactic and semantic rules against the word/phrase chunks to elevate the polarity of the word/phrase chunk to the entire expression includes the step of identifying a word/phrase chunk in the expression that toggles the polarity of the word/phrase chunk tagged with the polarity tag.

33. The method of claim 31, wherein the step of applying syntactic and semantic rules against the word/phrase chunks to elevate the polarity of the word/phrase chunk to the entire expression includes the step of performing with the assistance of a computer grammatical analysis on the expression.

34. The method of claim 31, further comprising the step of generating with the assistance of a computer the polarity lexicon from a domain-general lexicon that is tuned to a topical domain being explored.

35. The method of claim 34, further comprising the step of tuning with the assistance of a computer implemented graphical user interface the polarity lexicon by providing a user with candidates for addition, subtraction or exclusion to the polarity lexicon and for adding, subtracting or excluding each candidate from the polarity lexicon according to input received by the graphical user interface.

36. A computer implemented method for obtaining topical sentiments from a body of electronically stored communications comprising the steps of:

isolating with the assistance of a computer a first subset of the communications topical to a particular topic; and

isolating with the assistance of a computer a second subset of communications from the first subset of communications where the second subset of communications includes polar segments located in the first subset of communications.

37. The method of claim 36, wherein the polar segments are at least one of positive and negative.

38. The method of claim 36, wherein the polar segments located in the first subset of communications are positive and the method further includes the step of isolating a third subset of communications from the first subset where the third subset of communications include negative segments located in the first set of communications.

39. The method of claims 36 further comprising the step of associating the polar segments with the particular topic.

40. The method of claims 36 further comprising the step of displaying the polar segments on a computer display.

41. The method of claim 40, further comprising the step of displaying a communication from which the polar segment was extracted.

42. The method of claim 36, wherein the segment is at least one of a sentence, a phrase, a paragraph and an entire communication.

43. A computer implemented method for tuning a polarity lexicon for use in classifying polar expressions, comprising the steps of:

providing a polarity lexicon;

with the assistance of a computer implemented graphical user interface providing a user with candidate words for addition, subtraction or exclusion to the polarity lexicon; and

adding, subtracting or excluding each candidate word from the polarity lexicon according to input received by the graphical user interface.

44. The method of claim 43, wherein the step of providing a user with candidates for addition, subtraction or exclusion to the polarity lexicon includes a step of scanning a plurality of electronic messages collected for the topical domain for words that have the potential to be added to the lexicon.

45. The method of claim 44, wherein the scanning step includes a pattern based method that locates adjectives and adverbs that have a substantial chance of being polar.

46. The method of claim 44, wherein the scanning step locates candidate words by filtering the communication for words that appear at least a predefined number of times.

47. The method of claim 44, wherein the scanning step includes a pattern based method that locates adjectives and adverbs that have a substantial chance of being polar and locates candidate words by filtering the communication for words that appear at least a predefined number of times.

48. The method of claim 43 wherein the step of providing with a graphical user interface a user with candidate words for addition, subtraction or exclusion to the polarity lexicon includes the step of presenting each candidate word to the user with the word'"'"'s part of speech label and an example of that candidate word appearing in at least one electronic message collected for the topical domain.

49. A computer implemented method for obtaining topical sentiments from an electronically stored communication comprising the steps of:

determining with the assistance of a computer one or more topical expressions in the communication;

locating with the assistance of a computer one or more polar expressions in the communication; and

identifying an expression that is both a topical expression and a polar expression as containing a topical sentiment.

50. The computer implemented method of claim 49, wherein the steps of determining one or more topical expressions and locating one or more polar expressions are isolated steps performed on the same communication.

51. The computer implemented method of claim 50, wherein the step of determining one or more topical expressions includes a step of applying an automated text classifier on the communication and wherein the step of locating one or more polar expressions includes the step of utilizing a domain-specific lexicon and shallow NLP techniques.