1. Recognizing Stances in Online Debates Unsupervised opinion analysis method for debate-side classification. Mine the web to learn associations that are indicative of opinion stances. Formulate the debate side classification as an integer Linear Programming problem.

2. Recognizing Stances in Online Debates IPhone VS Blackberry. – Topic1 = iPhone, topic2 = Blackberry – Side1 = pro-iPhone, side2 = pro-Blackberry Give a post, analysis its side. – Post may express their opinion to some aspects of the topic. Some aspects are particular to one topic, while some aspects are shared. – Concession

3. Recognizing Stances in Online Debates Method: unsupervised approach – Finding opinions and pairing them with targets. Get opinion-target pairs. Find opinion: – Look up words in a subjectivity lexicon. All instances of those words are treated as opinions. – An opinion is assigned the prior polarity (positive (+), negative (-), neutral (*)). Pair with targets: – Build rule-based system to do syntactic analysis.

4. Recognizing Stances in Online Debates Syntactic rules:

5. Recognizing Stances in Online Debates Once opinion-target pairs are created, mask the identity of the opinion word, replacing the word with its polarity. – For instance, “pleasing interface” is converted to “interface+” (polarity-target pairs)

6. Recognizing Stances in Online Debates Learning aspects and preferences from web – Search web for pages which contain “iPhone” and “Blackberry” – Yield polarity-target pairs. – If the target in a polarity-target pair happens to be one of the topics, select the polarity-target pairs in its vicinity for further process. Vicinity is defined as the same sentence plus the following 5 sentences. (someone expresses an opinion about a topic, he or she is likely to follow it up with reasons for that opinion)

7. Recognizing Stances in Online Debates Learning aspects and preferences from web – where p = {+,−,*} and q = {+,−,* } denote the polarities of the target and the topic, respectively; j = {1, 2}; and i = {1...M}, where M is the number of unique targets in the corpus. For example, P(Mac+|interface+) is the probability that "interface" is the target of a positive opinion that is in the vicinity of a positive opinion toward "Mac."

8. Recognizing Stances in Online Debates Learning aspects and preferences from web

9. Recognizing Stances in Online Debates Debate-side classification – First, extract all polarity-target pairs in one post. – Second, Let N be the number of instances of polarity-target pairs in the post. For each instance Ij (j = {1...N}), we look up the learned probabilities of to create two scores, wj and uj :

10. Recognizing Stances in Online Debates Debate-side classification – Third, formulate the problem as an Integer Linear Programming problem. Maximum the objective function

11. Recognizing Stances in Online Debates Accounting for concession – Use the list of connectives from Concession and Contra- expectation. Like “while”, “nonetheless”, “however”. – If the connective is mid-sentence, the part of the sentence prior to the connective is considered conceded. – If the connective is sentence-initial, the sentence is split at the first comma. The first part is considered conceded – The opinion in the conceded part are interpreted in reverse. The weight corresponding to the sides wj and uj are interchanged.

12. Recognizing Stances in Online Debates Experiments – Baseline OpTopic systems. Only find all the topic1+, topic1-, topic2+ and topic2- instances in the post. OpPMI – Find opinion target pairs not only the topic, but also the words in the debated that are significantly related to either of the topic. – Calculate PMI between each noun in the post and topic over the whole web corpus. Assign each noun to the topic with higher PMI. – Next, the polarity-target pairs are found for the post, as before, and Equations 10 and 11 are used to assign a side to the post as in the OpTopic system, except that here, related nouns are also counted as instances of their associated topics.

13. Recognizing Stances in Online Debates Results – In our task, it is desirable to make a prediction for all the posts; hence #relevant = #Total posts. This results in Recall and Accuracy being the same. However, all of the systems do not classify a post if the post does not contain the information it needs. Thus, #guessed < #Total posts, and Precision is not the same as Accuracy.

14. Recognizing Stances in Online Debates Results

15. Recognizing Stances in Online Debates Regard the opinion word towards to the topic as cause? Regard vicinity polarity-target pairs as cause?

16. Classifying Latent User Attributes in Twitter Classify tweet users’ latent attributes, including gender, age, regional origin and political orientation. Supervised classification. Use below features for all four attributes classification. – Using lexical-feature based approaches – Extracting and utilizing sociolinguistics-inspired features.

17. Classifying Latent User Attributes in Twitter Attribute: – Gender: Male or Female – Age: below 30 or above 30 – Regional Origin: south India or north India – Political Orientation: Republican/Conservation vs. Liberal/Left/Democratic. Manually annotate training data.

18. Classifying Latent User Attributes in Twitter Network structure. (An initial investigation of potential features of classification) – The follower-following ratio: The ratio of number of followers of an user to the number of users he/she is following. – The follower frequency: The number of followers – The following frequency: The number of followees

19. Classifying Latent User Attributes in Twitter Network structure and communication behavior. (An initial investigation of potential features of classification)

20. Classifying Latent User Attributes in Twitter Network structure and communication behavior. (An initial investigation of potential features of classification)

21. Classifying Latent User Attributes in Twitter Communication behavior. (An initial investigation of potential features of classification) – Response frequency: percentage of tweets from the user that are replies – Retweet frequency: percentage of tweets that are retweets. – Tweet frequency: percentage of tweets that are from the user, uninitiated.

22. Classifying Latent User Attributes in Twitter Communication behavior. (An initial investigation of potential features of classification)

23. Classifying Latent User Attributes in Twitter Classification Models – Sociolinguistics feature models. Certain utterances like "umm", “uh-huh” are more prevalent among female speakers than their male counterparts.

24. Classifying Latent User Attributes in Twitter Classification Models – Sociolinguistics feature models. The templates from Table 2 resulted in 3774 unique instantiated feature types. The extracted features were used to learn an SVM based binary classier; we call this model socling.

25. Classifying Latent User Attributes in Twitter Classification Models – Ngram-features Model. Derive the unigrams and bigrams of the tweet text. Build another SVM-based classification model.

26. Classifying Latent User Attributes in Twitter Classification Models – Stacked Model. Finally, we employed a stacked model to do simple classifier stacking. We utilized another SVM for this task, but its features are the predictions from the Ngram- feature and Sociolinguistic models along with their prediction weights.

27. Classifying Latent User Attributes in Twitter Evaluation – Gender

28. Classifying Latent User Attributes in Twitter Evaluation – Age

29. Classifying Latent User Attributes in Twitter Evaluation – Regional Origin

30. Classifying Latent User Attributes in Twitter Evaluation – Political orientation

31. Classifying Latent User Attributes in Twitter Combine user attribute with sentiment or cause analysis??? – Different kind of people may have different opinion towards the same event.

32. Reference Recognizing Stances in Online Debates Classifying Latent User Attributes in Twitter