Domain-specific sentiment classification
A domain-specific sentiment classifier that can be used to score the polarity and magnitude of sentiment expressed by domain-specific documents is created. A domain-independent sentiment lexicon is established and a classifier uses the lexicon to score sentiment of domain-specific documents. Sets of high-sentiment documents having positive and negative polarities are identified. The n-grams within the high-sentiment documents are filtered to remove extremely common n-grams. The filtered n-grams are saved as a domain-specific sentiment lexicon and are used as features in a model. The model is trained using a set of training documents which may be manually or automatically labeled as to their overall sentiment to produce sentiment scores for the n-grams in the domain-specific sentiment lexicon. This lexicon is used by the domain-specific sentiment classifier.
1. A method of constructing a domain-specific sentiment classifier for classifying sentiment expressed by documents in a specified domain, comprising:
scoring sentiments expressed by a plurality of domain-specific documents responsive to a domain-independent sentiment lexicon;
identifying a set of high-sentiment documents responsive to the scoring;
creating a domain-specific sentiment lexicon based at least in part on the set of high-sentiment documents;
generating the domain-specific sentiment classifier responsive to the domain-specific sentiment lexicon; and
storing the domain-specific sentiment classifier.
2. The method of claim 1 , further comprising:
establishing the domain-independent sentiment lexicon by specifying a magnitude and polarity of sentiment expressed by each of a plurality of n-grams drawn from a domain-independent source.
3. The method of claim 1 , wherein the domain-independent sentiment lexicon specifies a magnitude and polarity of sentiment expressed by each of a plurality of n-grams and wherein scoring sentiment expressed by a plurality of domain-specific documents responsive to a domain-independent sentiment lexicon comprises:
for each domain-specific document:
determining whether the document includes an n-gram in the domain-independent sentiment lexicon;
calculating a score for the n-gram included in the document; and
calculating a sentiment score for the document responsive to the scores of any n-grams included within the document.
4. The method of claim 3 , wherein the score calculated for the n-gram included in the document is calculated responsive to one or more factors/techniques from the set consisting of:
a score for the n-gram specified in the domain-independent sentiment lexicon;
part-of-speech tagging responsive to a part-of-speech represented by the n-gram within the document;
detecting whether the n-gram is used in a negative manner within the document;
a location of the n-gram in the document; and
stemming to identify a root of the n-gram in the document.
5. The method of claim 1 , wherein identifying the set of high-sentiment documents responsive to the scoring comprises:
identifying a set of high-positive-sentiment documents; and
identifying a set of high-negative-sentiment documents.
6. The method of claim 1 , wherein creating the domain-specific sentiment lexicon based at least in part on the set of high-sentiment documents comprises:
identifying n-grams occurring within the documents in the set of high-sentiment documents;
filtering the identified n-grams to remove n-grams unlikely to express sentiment in the specific domain; and
storing the filtered n-grams as the domain-specific sentiment lexicon.
7. The method of claim 6 , wherein filtering the identified n-grams comprises:
removing extremely common n-grams.
8. The method of claim 1 , wherein generating the domain-specific sentiment classifier responsive to the domain-specific sentiment lexicon comprises:
building a model having n-grams of the domain-specific sentiment lexicon as features; and
training the model on a training corpus having the set of high-sentiment documents.
9. The method of claim 1 , wherein generating the domain-specific sentiment classifier responsive to the domain-specific sentiment lexicon comprises:
building a model having n-grams of the domain-specific sentiment lexicon as features; and
training the model on a training corpus having domain-specific documents having manually-labeled sentiment scores.
10. The method of claim 9 , wherein the model comprises a maximum entropy model.
11. The method of claim 9 , wherein the training generates sentiment scores for the n-grams of the domain-specific sentiment lexicon and further comprising:
storing the sentiment scores for the n-grams of the domain-specific sentiment lexicon with the domain-specific sentiment lexicon.
12. The method of claim 1 , further comprising:
using the domain-specific sentiment classifier to score a magnitude and polarity of sentiment expressed by a document in the specified domain.
13. A system for constructing a domain-specific sentiment classifier for classifying sentiment expressed by documents in a specified domain, comprising:
a computer-readable storage medium storing executable modules comprising:
a scoring module configured to score sentiments expressed by a plurality of domain-specific documents responsive to a domain-independent sentiment lexicon;
an analysis module configured to identify a set of high-sentiment documents responsive to the scoring;
a lexicon module configured to create a domain-specific sentiment lexicon based at least in part on the set of high-sentiment documents;
a classifier module configured to generate the domain-specific sentiment classifier responsive to the domain-specific sentiment lexicon; and
a storing module configured to store the domain-specific sentiment classifier; and
a processor configured to execute the modules stored by the computer-readable storage medium.
14. The system of claim 13 , wherein the domain-independent sentiment lexicon specifies a magnitude and polarity of sentiment expressed by each of a plurality of n-grams and wherein the scoring module is further configured to:
for each domain-specific document:
determine whether the document includes an n-gram in the domain-independent sentiment lexicon;
calculate a score for the n-gram included in the document; and
calculate a sentiment score for the document responsive to the scores of any n-grams included within the document.
15. The system of claim 13 , wherein the analysis module is further configured to:
identify a set of high-positive-sentiment documents; and
identify a set of high-negative-sentiment documents.
16. The system of claim 13 , wherein the lexicon module is further configured to:
identify n-grams occurring within the documents in the set of high-sentiment documents;
filter the identified n-grams to remove n-grams unlikely to express sentiment in the specific domain; and
store the filtered n-grams as the domain-specific sentiment lexicon.
17. The system of claim 16 , wherein filtering the identified n-grams comprises:
removing extremely common n-grams.
18. The system of claim 13 , wherein the classifier module is further configured to:
build a model having n-grams of the domain-specific sentiment lexicon as features; and
train the model on a training corpus having the set of high-sentiment documents.
19. The system of claim 13 , wherein the classifier module is further configured to:
build a model having n-grams of the domain-specific sentiment lexicon as features; and
train the model on a training corpus having domain-specific documents having manually-labeled sentiment scores.
20. The system of claim 19 , wherein the model comprises a maximum entropy model.
21. The system of claim 19 , wherein the training generates sentiment scores for the n-grams of the domain-specific sentiment lexicon and wherein the storing module is further configured to:
store the sentiment scores for the n-grams of the domain-specific sentiment lexicon with the domain-specific sentiment lexicon.
22. The system of claim 13 , wherein the scoring module is further configured to:
score a magnitude and polarity of sentiment expressed by a document in the specified domain using the domain-specific sentiment classifier.
23. A computer program product having a computer-readable storage medium having computer-executable code for constructing a domain-specific sentiment classifier for classifying sentiment expressed by documents in a specified domain, the computer-executable code comprising:
a scoring module configured to score sentiments expressed by a plurality of domain-specific documents responsive to a domain-independent sentiment lexicon;
an analysis module configured to identify a set of high-sentiment documents responsive to the scoring;
a lexicon module configured to create a domain-specific sentiment lexicon based at least in part on the set of high-sentiment documents;
a classifier module configured to generate the domain-specific sentiment classifier responsive to the domain-specific sentiment lexicon; and
a storing module configured to store the domain-specific sentiment classifier.
24. The computer program product of claim 23 , wherein the domain-independent sentiment lexicon specifies a magnitude and polarity of sentiment expressed by each of a plurality of n-grams and wherein the scoring module is further configured to:
for each domain-specific document:
determine whether the document includes an n-gram in the domain-independent sentiment lexicon;
calculate a score for the n-gram included in the document; and
calculate a sentiment score for the document responsive to the scores of any n-grams included within the document.
25. The computer program product of claim 23 , wherein the analysis module is further configured to:
identify a set of high-positive-sentiment documents; and
identify a set of high-negative-sentiment documents.
26. The computer program product of claim 23 , wherein the lexicon module is further configured to:
identify n-grams occurring within the documents in the set of high-sentiment documents;
filter the identified n-grams to remove n-grams unlikely to express sentiment in the specific domain; and
store the filtered n-grams as the domain-specific sentiment lexicon.
27. The computer program product of claim 26 , wherein filtering the identified n-grams comprises:
removing extremely common n-grams.
28. The computer program product of claim 23 , wherein the classifier module is further configured to:
build a model having n-grams of the domain-specific sentiment lexicon as features; and
train the model on a training corpus having the set of high-sentiment documents.
29. The computer program product of claim 23 , wherein the classifier module is further configured to:
build a model having n-grams of the domain-specific sentiment lexicon as features; and
train the model on a training corpus having domain-specific documents having manually-labeled sentiment scores.
30. The computer program product of claim 29 , wherein the model comprises a maximum entropy model.
31. The computer program product of claim 29 , wherein the training generates sentiment scores for the n-grams of the domain-specific sentiment lexicon and wherein the storing module is further configured to:
store the sentiment scores for the n-grams of the domain-specific sentiment lexicon with the domain-specific sentiment lexicon.
32. The computer program product of claim 23 , wherein the scoring module is further configured to:
score a magnitude and polarity of sentiment expressed by a document in the specified domain using the domain-specific sentiment classifier.