Twitter has brought much attention recently as a hot research topic in the domain of sentiment analysis. Training sentiment classifiers from tweets data often faces the data sparsity problem partly due to the large variety of short and irregular forms introduced to tweets because of the 140-character limit. In this work we propose using two different sets of features to alleviate the data sparseness problem. One is the semantic feature set where we extract semantically hidden concepts from tweets and then incorporate them into classifier training through interpolation. Another is the sentiment-topic feature set where we extract latent topics and the associated topic sentiment from tweets, then augment the original feature space with these sentiment-topics. Experimental results on the Stanford Twitter Sentiment Dataset show that both feature sets outperform the baseline model using unigrams only. Moreover, using semantic features rivals the previously reported best result. Using sentiment-topic features achieves 86.3% sentiment classification accuracy, which outperforms existing approaches.
"Subclassing and Composition – A Pythonic Tour of Trade-Offs", Hynek Schlawack
Alleviating Data Sparsity for Twitter Sentiment Analysis
1. Alleviating Data Sparsity For
Twitter Sentiment Analysis
Hassan Saif, Yulan He & Harith Alani
Knowledge Media Institute, The Open University, Milton
Keynes, United Kingdom
Making Sense of Microblogs – WWW2012 Conference
Lyon - France
2. Outline
• Hello World
• Motivation
• Related Work
• Semantic Features
• Topic-Sentiment Features
• Evaluation
• Demos
• The Future
3. Sentiment Analysis
“Sentiment analysis is the task of identifying
positive and negative opinions, emotions and
evaluations in text”
The main dish was The main dish was
It is a Syrian dish
delicious salty and horrible
Opinion Fact Opinion
3
4. Microblogging
• Service which allows subscribers to post short
updates online and broadcast them
• Answers the question: What are you doing
now?
• Twitter, Plurk, sfeed, Yammer, BlueTwi, etc.
4
7. Sense?
1400000
AGARWAL et al.
1200000 1143562
BARBOSA, L., AND FENG, J.
1000000
BIFET, A., AND FRANK, E.
800000
DIAKOPOULOS, N., AND SHAMMA, D. 713178
GO et al. 600000
He & Saif 400000
PAK & PAROUBEK 200000
And 0
1
Many Objective Tweets Subjective Tweets
Others
UK General Elections Corpus
8. Why
Private Sectors
Because It is Critical
df Keep In Touch
Public Sectors
10. Sentiment Analysis
Lexical Based Approach
Text Classification Problem
Right Features
Word Polarity
Building Better Dictionary
Machine Learning Approach
12. Twitter Sentiment Analysis
Related Work
• Distant Supervision
– Supervised classifiers trained from noisy labels
– Tweets messages are labeled using emoticons
– Data filtering process
Go et al., (2009) - Barbosa and Fengl. (2010) – Pak and Paroubek (2010)
13. Twitter Sentiment Analysis
Related Work
• Followers Graph & Label Propagation
– Twitter follower graph (users, tweets, unigrams
and hashtags)
– Start with small number of labeled tweets
– Applied label propagation method throughout the
graph.
Speriosu et al., (2009)
14. Twitter Sentiment Analysis
Related Work
• Feature Engineering
– Unigrams, bigrams, POS
– Microblogging features
• Hashtags
• Emoticons
• Abbreviations & Intensifiers
Agsrwal et al., (2011) – Kouloumpis et al (2011)
18. How!
Sentiment Topic Features
Extracts semantically hidden concepts
from tweets data and then
incorporates them into supervised
Extract latent topics and classifier training by interpolation
the associated topic
sentiment from the
tweets data which are
subsequently added into
the original feature space Semantic Features
for supervised classifier
training
19. Semantic Features
Shallow Semantic Method
Sport
@Stace_meister Ya, I have Rugby in an hour
Sushi time for fabulous Jesse's last day on dragons den Person
Dear eBay, if I win I owe you a total 580.63 bye paycheck
Company
21. Topic-Sentiment Features
Joint Sentiment Topic Model
JST1 is a four-layer generative model
which allows the detection of both
sentiment and topic simultaneously
from text.
The only supervision is word prior
polarity information which can be
obtained from MPQA subjectivity
lexicon.
Lin & He. 2009
22. Twitter Sentiment Corpus
Stanford University
Collected: the 6th of April & the 25th of June 2009
Training Set: 1.6 million tweets (Balanced)
Testing Set: 177 negative tweets & 182 positive tweets
http://twittersentiment.appspot.com/
23. Our Sentiment Corpus
• Training Set: 60K tweets
• Testing Set: 1000 tweets
• Annotated additional 640 tweets using
Tweenator
24. Evaluation
Extended Test Set
Method Accuracy
Unigrams 80.7%
Semantic replacement 76.3%
Semantic interpolation 84.0%
Sentiment-topic features 82.3%
Sentiment classification results on the 1000-tweets test set
25. Evaluation
Stanford Dataset
Method Accuracy
Unigrams 81.0%
Semantic replacement 77.3%
Sematic augmentation 80.45%
Semantic interpolation 84.1%
Sentiment-topic features 86.3%
(Go et al., 2009) 83%
(Speriosu et al., 2011) 84.7%
Sentiment classification results on the original Stanford Twitter Sentiment test set
28. Conclusion
• Twitter sentiment analysis faces data sparsity problem due to
some special characteristics of Twitter
• Semantic & topic-sentiment features reduce the sparsity
problem and increase the performance significantly
• Sentiment-topic features should be preferred over semantic
features for the sentiment classification task since it gives
much better results with far less features.
30
30. Future Work
Sentiment-Topic Model
Enhance Entity Extraction Methods
Attaching weight to extracted features
Semantic Smoothing Model
Statistical replacement
31. References
[1] AGARWAL, A., XIE, B., VOVSHA, I., RAMBOW, O., AND PASSONNEAU, R. Sentiment analysis of twitter data. In Proceedings of
the ACL 2011 Workshop on Languages in Social Media (2011), pp. 30–38.
[2] BARBOSA, L., AND FENG, J. Robust sentiment detection on twitter from biased and noisy data. In Proceedings of COLING
(2010), pp. 36–44.
[3] BIFET, A., AND FRANK, E. Sentiment knowledge discovery in twitter streaming data. In Discovery Science (2010), Springer, pp.
1–15.
[4] GO, A., BHAYANI, R., AND HUANG, L. Twitter sentiment classification using distant supervision. CS224N Project
Report, Stanford (2009).
[5] KOULOUMPIS, E., WILSON, T., AND MOORE, J. Twitter sentiment analysis: The good the bad and the omg! In Proceedings of
the ICWSM (2011).
[5]LIN, C., AND HE, Y. Joint sentiment/topic model for sentiment analysis. In Proceeding of the 18th ACM conference on
Information and knowledge management (2009), ACM, pp. 375–384.
[6] PAK, A., AND PAROUBEK, P. Twitter as a corpus for sentiment analysis and opinion mining. Proceedings of LREC 2010 (2010).
[7]SAIF, H., HE, Y., AND ALANI, H. Semantic Smoothing for Twitter Sentiment Analysis. In Proceeding of the 10th International
Semantic Web Conference (ISWC) (2011).
Can we extract sentiment from microblogs?Provide list of people who said users express more opinions on microblogscustomers are less controlled of their emotions when interacting in a social environment.So, yes we can do sentiment analysis over microblogs.
Now, Why do we need to do it?Millionsof status updates and tweets messages are shared everyday among usersFor public sectors: social data can be used to predict public political opinion which make easier to policy makers to put political strategies based on that.For private sector: Companies can track public opinion about their products and services
Two Main ApproachesLexical-based ApproachBuilding a better dictionaryFocuses on words and phrases as bearers of semantic orientationMachine Learning Approach Finding the Right FeaturesYet another form of text classification
To overcome the noisy label data
To overcome the noisy label data
To overcome the noisy label data
Previous work tried to overcome the sparsity problem partially and indirectly by either apply some data filtering or depending on different set of microblogs features. However,None of them studied the affect of data sparsity on the classification performance.
Why: many words and terms in the testing corpus doesn’t appear in the training corpus.
Compares the word frequency statistics of the tweets data we used in our experiments and the movie review dataX-axis shows the word frequency interval, e.g., words occur up to 10 times. (1-10), more than 10 times but up to 20 times (10-20), etc.Y-axis shows the percentage of words falls within certain word frequency interval.It can be observed that the tweets data are sparser than the movie review data since the former contain more infrequent words, with 93% of the words in the tweets data occurring less than 10 times (cf. 78% in the movie review data).