Sentiment-Analysis-of-Movies

How to run:

---

If running on console:

1.Save NBTrain.java and NBTest.java on disk . 2.Compile the java file by giving below command:

javac NBTrain.java javac NBTest.java

3. Run by giving the command java NBtrain java NBTest

Results are generated at Model file folder

Assumption: Train data should have two folders as "pos" and "neg" in train data folder. Dev data is also classified and present in two folders as per thier sentiment.

** The other projects on binary and bi-gram classifcation can also be compiled and executed in the same way.

---

If running on eclipse:

1. Import the HW6 package
2. Give the command line argument through Run->Configuration->Command Line Argument as
3. Run

---

Test Results:

Results on Dev Data:

• dev_result inside Result contain the classification of dev data Classifier_<MM-dd-yy_HH-mm-ss> file. Every row contains doc name, positive score, negative score, prediciton by classifier, given classification
• test_result inside Result contain the classification of test data as Classifier_<MM-dd-yy_HH-mm-ss> file Every row contains doc name, positive score, negative score, prediciton by classifier
• Model.txt is the model file generated on train data The first line of prediction file contains the metadata. Next each row contains terms, term frequency in pos class and term freq in neg class
• PosToNeg_<MM-dd-yy_HH-mm-ss> file contains the terms soretd by positive to negative log ratio of term weights this contains the term, ratio, pos frequency and neg frequency
• NegToPos_<MM-dd-yy_HH-mm-ss> file contains the terms soretd by negative to positive log ratio of term weights this contains the term, ratio, pos frequency and neg frequency
• Top20PosToNeg file contains the top 20 words sorted on positive to negative log ratio of term weights this contains the term, ratio, pos frequency and neg frequency
• Top20NegToPos file contains the top 20 words sorted on negative to positive log ratio of term weights this contains the term, ratio, pos frequency and neg frequency

Percentage of correct classifcation in dev data wrt to classifer prediction: Positive Percentage Classification: 73% Negative Percentage Classification: 84%

Classification of test data: Positives: 86 Negatives: 114

---

Model Modification

---

These solutions are stored in "Modified Version" Folder

For modified version we used two solutions:

1. Using Binary relevance with Laplace smoothing
2. Using Bi-gram Model with Laplace Smoothing

Proposed Modification: We used Binary Bayesian Classifier instead of multinomial classifier and found that terms to which it assigns highest probality scores do signify the positive and negative sentiments. Similary we used Bi-gram model but found that it tends to align towards negative probablity , also the top words predicted for positive and negative words do not specify and sentiments as the terms have less occurances as a combination across collection or they form a term with a punctuation which reduces the relevance of the word.

The comparision between each of the solution are described below on dev data:

                       Neg%        Pos%            Top Negative Words               	Top Positive Words

Multinomial Classifier: 84 73 [ &nbsp,jolie,seagal,brenner ] [shrek,mulan,gattaca,flynt]

Binary Classifier : 57 95 [ Bad, worst , boring, stupid] [life,both,world,great,perfect,best,performance]

Bi-gram Classifier : 99 55 [  ,.&nbsp, 'this mess', mars "] [black cauldron , drunken master, "chicken, story2]

Also the classification by the binary classifier and bi-gram classifier are as below:

Binary classifier: Positives: 143 Negatives: 57

Bi-gram classifier: Positives: 47 Negatives: 153

After analysing we found that Binary Classifier tends to outperform n-grams classifier if the language used is straightforward and use proper sentiment terms. For example the below sentence will be classified as positive although its negative by binary classifier. For ex: I was expecting the movies to be great and awesome but it was not so good. Also the results shows that Multinomial Bayes Classification contains highest weights of many terms which does not signify any sentiments such as &nbsp. This polarity is due to frequency associated with each terms.