1. Peiti Li 1, Shan Wu 2, Xiaoli Chen 1 1 Computer Science Dept. 2 Statistics Dept. Columbia University 116th Street and Broadway, New York, NY 10027, USA introducing Movie Review

2. It is a fast and more direct way for people to share their opinions on a topic Why ?

3. Python Twitter Search API + Stream API

4. Opinion Mining or Sentiment Analysis Computational study of opinions, sentiments, subjectivity, attitudes

5. Just like a text classification task but different from topic-based text classification In topic-based text classification (e.g., computer, sport, science), topic words are important. But in sentiment classification, opinion/sentiment words are more important, e.g., awesome, great, excellent, horrible, bad, worst, etc.

6. Structure the unstructured: Natural language text is often regarded as unstructured data Besides data mining, we need NLP technologies Why a HARD task? I bought an iPhone a few days ago. It is such a nice phone. The touch screen is really cool. The voice quality is clear too. It is much better than my old Blackberry, which was a terrible phone and so difficult to type with its tiny keys. However, my mother was mad with me as I did not tell her before I bought the phone. She also thought the phone was too expensive,… Credits: Bing Liu for this example

7. Tell people whether to go to buy a movie ticket using tweets Classify the tweet as either positive or negative Give a rating of the movie based on tweets

8. Different Machine Learning Approaches Accuracies Table from: Bo Pang et al. 2002. Thumbs up? Sentiment Classification using Machine Learning Techniques. In Proc. Of the ACL, pp. 79-86. Association for Computational Linguistics

9. Our approach is Naïve Bayes P(sentiment | sentence) = P(sentiment)P(sentence | sentiment) / P(sentence) Smoothing: P(token | sentiment) = (count(this token in class) + 1) / (count(all tokens in class) + count(all tokens)) We didn’t use any third-party classifier, we coded our classifier all by ourselves. Reason: want to explore what is under the hook; tune the algorithm structure according to the experiment result

10. Getting Started.

11. » Dev set: The movie review dataset provided by Bo Pang and Lillian Lee, Cornell University sentence_polarity_dataset_v1.0 5331 positive, 5331 negative » Real set: Tweets about a specific movie Cannot tell exact number Twitter Search API(REST): last 6-7 days Twitter Stream API: real timeline (Drawbacks: REST API has rate limiting; Stream data takes time to collect.) Dataset

12. Top 100 words including stopwords

13. Better and better but…. Baseline model is the Naïve Bayes, without any nontrivial text preprocessing; punctuations excluded, stopwords included Tuned model still Naïve Bayes, better feature extraction technique: eliminating low information features. Best unigram model, best unigram and bigram model

14. Dev set result: Trainset 5000, Testset 331RecallSpecificityAccuracy Baseline76.13%82.78%79.46% Baseline, stopwords removed 75.83%79.46%77.64% Best unigram, stopwords not removed 83.99%85.20%84.60% Best unigram, stopwords removed 82.78%85.80%84.29% Best unigram and bigram, stop words not removed N/A 78.24% Takes 1 hour! Intel Core i5 laptop died in the middle because of too hot for too long Observation: definitely not consider bigrams, but still don’t know whether we should remove the stopwords

15. 5 neg, 87 pos 150 tweets 75 labeled by Xiaoli, 75 labeled by Shan 75 labeled by Xiaoli, 75 labeled by Shan 150 tweets 76 neg, 32 pos

16. Regular expression 1: (?:@\S*|#\S*|http(?=.*://)\S*) Regular expression 2: (#[A-Za-z0-9]+) | (@[A-Za-z0-9]+)|([^0-9A-Za-z \t])|(\w+:\/\/\S+) (All punctuations removed) HugoMuppetstogether stopwords remv64.13%64.81%64.50% stopword incld63.04%54.63%58.50% stopwords remv70.65%62.96%66.5% stopwords incld65.22%53.70%59.00% Results on the 2 recent movies(Real set) Which regular expression should we choose based on this result? Hard to say…. :-(

17. . lingPipe, Twendz, Twitter Sentiment, tweetfeel We moved our attention to:

18. twittersentiment.appspot.com They are new too.

19. www.tweetfeel.com Our classifier get the exact same results with them, but wait…

20. Two pieces of tweet made us frown :-(

21. Emoticons play a role!!! >:] :-) :) :o) :] :3 :c) :> =] 8) =) :} :^) >:D :-D :D 8-D 8D x-D xD X-D XD =-D =D =-3 =3 :P FTW :'( ;*( :_( T.T T_T Y.Y Y_Y >:[ :-( :( :-c :c :-< :< :-[ :[ :{ >.>. :\ >:/ :-/ :-. :/ :\ =/ =\ :S

22. So we choose the regular expression that will keep emoticons And we build a dictionary to eliminate all the punctuations that appear alone '`','~','!','@','#','$','%','^','&','*','(',')','- ','_','+','=','{','}','[',']',';',':','"',"'",' ',',','.','?','|','\\','/'

23. Finally, the python begins to catch the twittering bird…….. Demo

24. “Happy” Feet? So all tweets are positive? We still need to do more semi- supervised learning. 1.Specific bigrams like “don’t love” 2.Finer classifier which can exclude objectives 3. Detect and remove annoying movie name like “Happy Feet” 4. Give more weights to dominant words like “excellent”, “worst” 5. Our final task: Give ratings

25. Thank you all! Thank you STAT 4240! Thank you Columbia!