Smart Sentiment Classifier for Product Reviews

• US 20080249764A1
• Filed: 12/05/2007
• Published: 10/09/2008
• Est. Priority Date: 03/01/2007
• Status: Abandoned Application

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Huang, Shen, Chen, Zheng, Cao, Yunbo, Sun, Jian-Tao, Zhou, Ming, Wang, Jian, Lin, Chin-Yew, Ponath, Christoph R., Bao, Ling
• Application Number:
  US11/950,512
• Publication Number:
  US 20080249764A1
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G06F 40/30 Semantic analysis

Abstract

A sentiment classifier is described. In one implementation, a system applies both full text and complex feature analyses to sentences of a product review. Each analysis is weighted prior to linear combination into a final sentiment prediction. A full text model and a complex features model can be trained separately offline to support online full text analysis and complex features analysis. Complex features include opinion indicators, negation patterns, sentiment-specific sections of the product review, user ratings, sequence of text chunks, and sentence types and lengths. A Conditional Random Field (CRF) framework provides enhanced sentiment classification for each segment of a complex sentence to enhance sentiment prediction.

Claims

1. A method, comprising:

applying a text analysis to a received text to determine a first sentiment classification;

applying a complex features analysis to the received text to determine a second sentiment classification; and

combining the first and second sentiment classifications to achieve a sentiment prediction for the received text.

2. The method as recited in claim 1, wherein combining the first and second sentiment classifications includes:

weighting the first sentiment classification according to a confidence score associated with the text analysis and weighting the second sentiment classification according to a confidence score associated with the complex features analysis; and

linearly combining the weighted first sentiment classification and the weighted second sentiment classification to achieve the sentiment prediction.

3. The method as recited in claim 1, wherein the text analysis comprises an analysis of full-text text information, including determining a sequence of terms in sentences of the received text.

4. The method as recited in claim 1, wherein the complex features analysis comprises an analysis of opinion-carrying words in the received text, user rating information associated with the received text, sentiments associated with sections of the received text, negation words and patterns in the received text, and sentence types in the received text.

5. The method as recited in claim 4, further comprising extracting the opinion-carrying words, including:

tagging sentences with positive and negative sentiments with part-of-speech information, wherein N-grams (1≦

n<

5) are extracted;

filtering unigrams and associated part-of-speech information, wherein only unigrams with adjective, verb, adverb, or noun tags qualify as opinion-carrying word candidates;

assigning a cross entropy score and a Chi-square score to each candidate opinion-carrying word;

calculating a similarity of each opinion-carrying word candidate with pre-selected seed opinion words according to the equation

dist(w_i,S_c)=max {sim(w_i,p),pε

S_c},cε

{pos,neg};

determining a score for each opinion-carrying word candidate using cross entropy score and/or Chi-square score and the calculated similarity; and

determining a set of opinion-carrying words by ranking the scores.

6. The method as recited in claim 1, further comprising separately training a full-text sentiment classification model and a complex features sentiment classification model to support the text analysis and the complex features analysis.

7. The method as recited in claim 6, wherein the full-text sentiment classification model comprises a trigram-based Naive Bayesian model.

8. The method as recited in claim 6, wherein separately training the full-text sentiment classification model and the complex features sentiment classification model includes analyzing training data that includes sentences that have associated sentiment classifications assigned.

9. The method as recited in claim 6, wherein training the full-text sentiment classification model and training the complex features sentiment classification model are performed offline and processing the received text to achieve the sentiment prediction is performed online.

10. The method as recited in claim 6, further comprising associating a confidence score or a confidence rating with the sentiment prediction.

11. The method as recited in claim 6, further comprising training the full-text sentiment classification model and training the complex features sentiment classification model from different feature sets.

12. The method as recited in claim 1, further comprising segmenting sentences of the received text into chunks of words and constructing opinion classification features using both sentence information and sequential information of the chunks.

13. The method as recited in claim 12, wherein constructing opinion classification features includes modeling the text chunks of a sentence using a Conditional Random Field (CRF) framework.

14. The method as recited in claim 12, wherein if a sentence of the received text includes an indicator word, then splitting the sentence into chunks at the indicator word and assigning a sentiment orientation to each chunk and an overall sentiment orientation to the entire sentence, wherein the indicator word is selected from the group of indicator words consisting of “

but,”

“

if,”

“

however,” and

“

although.”

15. The method as recited in claim 1, wherein the sentiment classifications are selected from the group of sentiment classifications consisting of “

positive,”

“

negative,”

“

mixed,”

“

neutral,” and

“

none.”

16. A system, comprising:

a full text analyzer to provide a first sentiment classification of a received text;

a complex features analyzer to provide a second sentiment classification of the received text; and

an ensemble classifier to combine the first sentiment classification and the second sentiment classification into a sentiment prediction for the received text.

17. The system as recited in claim 16, further comprising:

a full text sentiment classification model for modeling sentiment associated with a sequence of terms in sentences of the received text;

a complex features sentiment classification model for modeling sentiment associated with non-text features of the received text, wherein the non-text features include one of an opinion feature, a negation word feature, a negation word pattern, a section of the product review with an associated sentiment, a user review rating, a type of sentence used to express a user opinion, a sequence of text chunks with respective sentiments, and a sentence length; and

wherein the full text sentiment classification model and the complex features sentiment classification model are trained separately.

18. The system as recited in claim 16, wherein the ensemble classifier assigns weights to the first sentiment classification and the second sentiment classification and executes a linear combination of the weighted first sentiment classification and the weighted second sentiment classification to provide the sentiment prediction.

19. The system as recited in claim 16, further comprising a chunk Conditional Random Field (CRF) framework for segmenting sentences of the received text into chunks and training a CRF model to predict a category of sentiment orientation for each chunk based on a set of training sentences.

20. An ensemble sentiment classifier for sentiment analysis of a product review, comprising:

means for applying a full-text analysis to a sentence of the product review based on a full text sentiment model trained from a first set of product review features;

means for applying a complex features analysis to the sentence based on a complex features sentiment model trained from a second set of product review features; and

means for weighting and combining the full-text analysis and the complex features analysis into a sentiment prediction for each sentence of the product review.