Method and system for analysing sentiments

• US 10,402,745 B2
• Filed: 09/30/2013
• Issued: 09/03/2019
• Est. Priority Date: 09/28/2012
• Status: Active Grant

• Current Assignee: 9417-2129 Quebec Incorporated
• Original Assignee: Semeon Analytics, Inc. (ATW Tech, Inc.)
• Inventors: Papadopoullos, Alkis, Desbiens, Jocelyn
• Primary Examiner(s): Chen, Alan
• Application Number:
  US14/432,436
• Publication Number:
  US 20150286953A1
• Time in Patent Office: 2,164 Days
• Field of Search: None
• CPC Class Codes: G06F 40/30 Semantic analysis
• CPC Class Codes: G06N 20/00 Machine learning
• CPC Class Codes: G06N 5/022 Knowledge engineering; Knowledge acquisition
• CPC Class Codes: G06N 7/023 Learning or tuning the parameters of a fuzzy system

Abstract

The present system and method for analyzing sentiment is based on fuzzy set theory and clustering to classify text as positive, negative, or objective. The method for training and testing a document collection to analyze sentiment comprises computing a frequency matrix comprising at least one row and at least column vectors, executing term reduction of the terms, enumerating categories, computing centroids of the enumerated categories and computing a fuzzy polarity map. The row vectors may correspond to terms and the column vectors may correspond to documents. The frequencies of terms in a document indicate the relevance of the document to a query.

Claims

1. A computer-implemented method for training a document collection to analyze sentiment comprising:

a) receiving a text expression from a computerized device;

b) a processor computing a frequency matrix from the received text expression, the frequency matrix comprising at least one row and at least column vectors, each row representing at least one term in a vocabulary each column representing at least one document of the document collection;

c) the processor executing a term reduction of the terms to remove useless terms in the frequency matrix;

d) the processor enumerating a plurality of categories of the document collection, the categories comprising at least 3 sentiment classes and at least 8 gradations for each sentiment classes, the classes comprising positive, negative, and neutral sentiments;

e) the processor computing centroids of the enumerated categories; and

f) the processor computing a fuzzy polarity map by;

i) building a polarity histogram of all terms belonging to the inputted expression;

ii) computing skewness of the polarity histogram;

iii) categorizing the inputted expression in terms of the computed skewness.

2. A method for training a document collection to analyze sentiment as claimed in claim 1, wherein the gradations of positive class comprises at least very positive, positive, lightly positive, the gradations of negative class comprises at least very negative, negative, or lightly negative and the gradation of neutral comprises balanced and true neutral.

3. A method for training a document collection to analyze sentiment as claimed in claim 1, wherein the frequencies of terms in a document are relative frequencies based on document length normalization, weighted based on information theory metrics adapted to assign a frequency score to the term based on normalized tf/idf and a proxy for an information index.

4. A method for training a document collection to analyze sentiment as claimed in claim 2, wherein every term of the frequency matrix is conceived as a separate dimension of a vector space and wherein a document is a formal vector.

5. A method for training a document collection to analyze sentiment as claimed in claim 1, wherein the distance between the vectors is measured by a geometrically-motivated function.

6. A method for training a document collection to analyze sentiment as claimed in claim 5, wherein the geometrically motivated function is one of the following functions:

L₁ norm, L₂ norm and cosine function.

7. A method for training a document collection to analyze sentiment as claimed in claim 1, wherein the step to execute term reduction comprises an initial step to apply one or more linguistic pre-treatments.

8. A method for training a document collection to analyze sentiment as claimed in claim 7, wherein the first linguistic pre-treatment is based on the extraction of all linguistic units used for document representation.

9. A method for training a document collection to analyze sentiment as claimed in claim 8, wherein the step to apply several linguistic pre-treatments is done through chained filters.

10. A method for training a document collection to analyze sentiment as claimed in claim 1, wherein the step to enumerate categories uses one or more training collections which are divided in a plurality of categories based on the sentiment type analysis to be executed.

11. A method for training a document collection to analyze sentiment as claimed in claim 10, wherein a polarity detection requires three categories:

positive, negative and objective, and wherein a subjectivity detection requires the three other categories;

subjective, objective and other.

12. A method for training a document collection to analyze sentiment as claimed in claim 10, wherein the step to compute category centroids assigns a cluster to each enumerated category.

13. A method for training a document collection to analyze sentiment as claimed in claim 1, wherein the step to enumerate categories uses an unsupervised clustering method.

14. A method for training a document collection to analyze sentiment as claimed in claim 13, wherein the unsupervised clustering method is the Spherical k-means method.

15. A method for training a document collection to analyze sentiment as claimed in claim 14, wherein clusters are generated and are used as an input to the Spherical k-means method.

16. A method for training a document collection to analyze sentiment as claimed in claim 1, further comprising computing text subjectivity of the inputted expression by:

a) building the polarity histogram of all terms belonging to the inputted expression;

b) computing kurtosis of the polarity histogram;

c) categorizing inputted expression in terms of the computed kurtosis.