Veeranna 'A 'Y

1
Sentiment Analysis

• Veeranna .A .Y
• 05005023

2
Hindi SentiWordNet

• Giving sentiment scores to every Hindi synset
  just like English SentiWordNet.
• Three scores pos, neg obj scores to every
  synset.
• Claim Sentiment scores given to English word
  senses also apply to their Hindi counterpart.
• Hindi-English synset linking work is in progress,
  half of Hindi WordNet has been mapped to English
  WordNet.

3
Building Hindi SentiWordNet

• Various mappings required
• Hindi-English synset mapping (mapped to English
  WordNet ver-2.1 is used)
• ver-2.1 to ver-2.0 synset mapping (SentiWordNet
  is based on ver-2.0)
• ver-2.0 to SentiWordNet sense no's mapping
• Sense no's are not same across all the vers of
  English WordNet, hence mapping of synsets across
  WordNet vers is required.
• ver-2.0 SentiWordNet offset no's doesnt match.
• Currently 12408 sense have been linked given
  senti-scores.

4
Table 1 Difference in Sense no's across WordNet
ver 2.1 , 2.0 SentiWordNet
Table 2 Examples from Hindi SentiWordNet
5
Connotation

• Connotation is a subjective cultural and/or
  emotional coloration addition to the explicit or
  denotative meaning of any specific word or phrase
  in a language, i.e. emotional association with a
  word.
• Connotations are used to express opinions.
• E.g. sea of people, you are a dog, stock
  market hit a decade low.

6
Feature Engineering

• Connotation of word can give a sentence positive,
  negative objective interpretation.
• E.g. you are a dog has a negative
  interpretation of a person.
• every child needs a good home has a positive
  interpretation of home.
• How much can a word shift the interpretation of
  sentence to either positive or negative from
  neutral can be measured by Interpretability score
  Dan Tufis (FASSBL,2008)

7
Feature Engineering

• where, maxP(swk) and maxN(swk) representing the
  highest positive and negative scores among the
  senses of swk senti-word. 0 IS 1.
• Currently in SentiWordNet, the senti-words with
  the highest interpretability score (IS 0.875)
  are pretty, immoral and gross.
• Words with IS score IS gt n (n is a parameter)
  are chosen as feature words for experiment.
• Each document is represented as a vector of these
  features.

8
Naïve Bayes

• Finding the probability of class being c given
  that the document is d .
• The document d is classified as class
• Each document d is represented by the document
  vector
• 
  , where ni(d)be the number of times fi occurs in
  document d.
• Document vector can be made in two ways, by
  letting ni(d) to take either the count of
  features or take the binary value (0 or 1)
  depending on the presence or absence of feature.

9
Results

• Experiment was done using 1000 positive 1000
  negative documents.
• Testing was done with unigrams only with
  unigrams multi-grams.
• It was observed that as the no of features
  decreased accuracy increased.
• Naive Bayes gives high accuracy when document
  vector is taken as frequency count than when
  taken as binary presence value
• Highest accuracy of 90.2 was achieved with
  features being Unigrams multi-grams ni(d) set
  to frequency count. IS 0.5 (for Unigrams 500
  unigrams) IS 0.2 (for multi-grams 1740
  multi-grams.

10
SVM

• Idea behind the training procedure is to find a
  hyperplane, represented by vector w , that not
  only separates the document vectors (dj) in one
  class (cj) from those in the other, but for which
  the separation, or margin, is as large as
  possible
• where the aj s are obtained by solving a
  dual optimization problem.
• package was used for training and
  testing, with all parameters set to their default
  values.

11
Results

• Experiment was done using 1000 positive 1000
  negative documents.
• Testing was done with unigrams only with
  unigrams multi-grams.
• SVM gives good accuracy when document vector is
  taken as binary presence value than when taken as
  frequency count.
• Accuracy of 73.35 was achieved with features
  being Unigrams multi-grams ni(d) set to
  frequency count. IS 0.5 (for Unigrams 500
  unigrams) IS 0.1 (for multi-grams 7215
  multi-grams).
• Accuracy of 74.55 was achieved with features
  being Unigrams multi-grams ni(d) set to
  binary presence value. IS 0.5 (for Unigrams
  500 unigrams) IS 0.2 (for multi-grams 1740
  multi-grams.

12
Minimum Graph Cut

• xis are sentences in a document.
• Framework
• indj (xi) estimate that xi belongs to Cj based
  on features present in the sentence only
  (individual score).
• assoc (xi, xk) estimate that xi xk belong to
  the same class (association score).
• Finding minimum partition cost problem reduces to
  finding minimum cut in graph.
• Minimum Cut in a graph finding a cut through
  edges in a graph such that partitioned components
  are connected graphs and sum of edges lying on
  cut should be minimum.

13

• Fig Graph cut based subjectivity extraction
• Courtesy Lillian Lee, et al. ACL 2004.

14

• Fig Polarity classification via subjectivity
  detection
• Courtesy Lillian Lee, et al. ACL 2004

15
Results

• Experiment was done using 1000 positive 1000
  negative documents, 5000 subjective 5000
  objective sentences.
• Either Naive Bayes or SVM along with graph cut
  method can be used as Subjectivity detector at
  the initial phase to get an subjective extract of
  documents.
• Similarly both can be used as polarity detector
  to classify the subjective extract got in initial
  phase.
• Accuracy of 92.4 was achieved with Naive Bayes
  as subjectivity detector graph cut, to get
  subjective extract again Naive Bayes as
  polarity detector to get final classification.
  Other parameters were, features being Unigrams
  multi-grams ni(d) set to frequency count, IS
  0.5 (for Unigrams 500 unigrams) IS 0.2 (for
  multi-grams 1740 multi-grams.

16
Comparison with existing algorithms without
SentiWordNet

• Table 3 Comparison of accuracies of existing
  methods with without using SentiWordNet
• Comparison is done only for best accuracies
  achieved in both cases. And also parameters vary
  in both cases.
• There is an increase in accuracy by including
  SentiWordNet into current existing algorithms for
  Sentiment Analysis.
• No of features have been reduced from previous
  methods.

17
Conclusion

• Direct mapping of English Sentiment scores to
  Hindi Synsets is justified when the mapping of
  Hindi-English Synsets mapping of synset nos
  across versions of WordNet is justified.
• Main aim of the SentiWordNet is to capture the
  intricacies of texts. Hence SentiWordNet gives
  sentiment scores to each synset rather than just
  hard binary classification of synsets. One of the
  intricacy of texts is Connotation which has been
  captured by giving each synset an
  Interpretability Score.
• Sentiment Analysis based on Connotation Analysis
  shows good results. Connotations are good at
  capturing Positive, Negative objective
  interpretations of texts.
• There is an increase in accuracy by including
  SentiWordNet connotation into current existing
  algorithms for Sentiment Analysis.

18
References

• Pang Bo, Lillian Lee, and Shivakumar
  Vaithyanathan. Thumbs up? Sentiment
  Classification using Machine Learning Techniques.
  In EMNLP 2002.
• Pang Bo, Lillian Lee. A Sentimental Education
  Sentiment Analysis Using Subjectivity
  Summarization Based on Minimum Cuts. In
  Proceedings of ACL 2004.
• "CONNOTATION ANALYSIS" Dan Tufis (FASSBL,2008)
• J.Ramanand, Akshay Ukey, Brahm Kiran Singh and
  Pushpak Bhattacharyya, Mapping and Structural
  Analysis of Multilingual WordNet, IEEE Data
  Engineering Bulletin, 30(1), March 2007.

19

• Thank You
• Questions?