swapna somasundaran [email protected]

60
Politics and Social media: The Political Blogosphere and the 2004 U.S. election: Divided They Blog Crystal: Analyzing Predictive Opinions on the Web Swapna Somasundaran [email protected]

Upload: darryl

Post on 19-Jan-2016

48 views

Category:

Documents


0 download

DESCRIPTION

Politics and Social media: The Political Blogosphere and the 2004 U.S. election: Divided They Blog Crystal: Analyzing Predictive Opinions on the Web. Swapna Somasundaran [email protected]. The Political Blogosphere and the 2004 U.S. election: Divided They Blog Link based Approach - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Swapna Somasundaran swapna@cs.pitt

Politics and Social media:

The Political Blogosphere and the 2004 U.S. election: Divided They Blog

Crystal: Analyzing Predictive Opinions on the Web

Swapna Somasundaran

[email protected]

Page 2: Swapna Somasundaran swapna@cs.pitt

2

Politics and Social media

The Political Blogosphere and the 2004 U.S. election: Divided They Blog

• Link based Approach

• Studies linking patterns between blogs just before the presidential elections

Crystal: Analyzing Predictive Opinions on the Web

• Language based approach

• Uses Linguistic expression of opinion to predict election results

Page 3: Swapna Somasundaran swapna@cs.pitt

3

The Political Blogosphere and the 2004 U.S. election: Divided They

Blog

Lada A. Adamic, Natalie Glance

Page 4: Swapna Somasundaran swapna@cs.pitt

4

Motivation: Social media and Politics

2004:• Harnessing grass root support

– Howard Dean’s campaign

• Breaking stories first – Anti-Kerry video

2007:

Page 5: Swapna Somasundaran swapna@cs.pitt

5

Page 6: Swapna Somasundaran swapna@cs.pitt

6

Page 7: Swapna Somasundaran swapna@cs.pitt

7

Outline

• Data collection

• Analysis

• Conclusions

• Similar work

Page 8: Swapna Somasundaran swapna@cs.pitt

8

Data

Web log directories_________________________

Web log directories_________________________

Page 9: Swapna Somasundaran swapna@cs.pitt

9

DataConservative blogs

Conservative blogs

Web log directories_________________________

Web log directories_________________________ Liberal blogsLiberal blogs

Page 10: Swapna Somasundaran swapna@cs.pitt

10

DataConservative blogs

Conservative blogs

Web log directories_________________________

Web log directories_________________________ Liberal blogsLiberal blogs

blogblog

Page 11: Swapna Somasundaran swapna@cs.pitt

11

DataConservative blogs

Conservative blogs

Web log directories_________________________

Web log directories_________________________ Liberal blogsLiberal blogs

blogblog

Page 12: Swapna Somasundaran swapna@cs.pitt

12

DataConservative blogs

Conservative blogs

Web log directories_________________________

Web log directories_________________________ Liberal blogsLiberal blogs

blogblog

1494 Blogs

Page 13: Swapna Somasundaran swapna@cs.pitt

13

Citation network

blogblog

Page 14: Swapna Somasundaran swapna@cs.pitt

14

Citation network

blogblogblogblog

blogblog

blogblogblogblog

Page 15: Swapna Somasundaran swapna@cs.pitt

15

Analysis: Citation network

Page 16: Swapna Somasundaran swapna@cs.pitt

16

Analysis: Citation network

91%

Page 17: Swapna Somasundaran swapna@cs.pitt

17

Analysis: Citation network

Conservative Blogs show a greater tendency to link

Page 18: Swapna Somasundaran swapna@cs.pitt

18

Analysis: Citation network84%

82%

74%

67%

Conservative Blogs show a greater tendency to link

Page 19: Swapna Somasundaran swapna@cs.pitt

19

Analysis: Posts

Data :

• Top 20 blogs from each each category

• Extract posts from these for a span of 2.5 months.

• 12470 left leaning, 10414 right leaning posts.

Page 20: Swapna Somasundaran swapna@cs.pitt

20

Analysis: Strength of community# of posts in which

one blog cited another blog

Remove links if fewer than 5

citations

Remove links if fewer than 25

citations

Page 21: Swapna Somasundaran swapna@cs.pitt

21

Analysis: Strength of community

Right-leaning blogs have denser structure of strong connections

than the left

Page 22: Swapna Somasundaran swapna@cs.pitt

22

Analysis: Interaction with mainstream media

Links to news articles

Page 23: Swapna Somasundaran swapna@cs.pitt

23

Analysis: response to CBS news item

Page 24: Swapna Somasundaran swapna@cs.pitt

24

Analysis: Occurrences of names of political figures

Page 25: Swapna Somasundaran swapna@cs.pitt

25

Analysis: Occurrences of names of political figures

Left leaning bloggers spoke more about Republicans and vice versa

People support their positions by criticizing those of the political figures they dislike

Page 26: Swapna Somasundaran swapna@cs.pitt

26

Conclusions

• Clear division of blogosphere– Links– Topics and people

• Conservative blogs are more likely to link.

Page 27: Swapna Somasundaran swapna@cs.pitt

27

Future work/ Extensions

• Include more blogger types

• Single/multi author distinction

• Spread of topics due to network structure

• …?

Page 28: Swapna Somasundaran swapna@cs.pitt

28

Some Similar Work

• Political Hyperlinking in South Korea: Technical Indicators of Ideology and Content, Park et al. Sociological Research Online, Volume 10, Issue 3, 2005

• Weblog Campaigning in the German Bundestag Election 2005 , Albrecht et al., ,Social Science Computer Review , Volume 25 ,  Issue 4 ,November 2007

• Friends, foes, and fringe: norms and structure in political discussion networks, Kelly et al., International conference on Digital government research , 2006

• 1000 Little Election Campaigns:Utilization and Acceptance of Weblogs in the Run-up to the German General Election 2005 Roland Abold, ECPR Joint Session., Workshop 9: ‘Competitors to Parties in Electoral Politics, 2006

Page 29: Swapna Somasundaran swapna@cs.pitt

29

Some interesting links

• http://www.politicaltrends.info/poltrends/poltrends.php

– political trend tracker - tracks sentiments in political blogs, and reports daily statistics

Page 30: Swapna Somasundaran swapna@cs.pitt

30

Page 31: Swapna Somasundaran swapna@cs.pitt

31

Page 32: Swapna Somasundaran swapna@cs.pitt

32

Some interesting links:

• Visualization of the blogosphere during French elections– http://www.observatoire-presidentielle.fr/?pageid=3

– http://www.fr2007.com/?page_id=2

Page 33: Swapna Somasundaran swapna@cs.pitt

33

Some Interesting Links:

• Political wiki:– http://campaigns.wikia.com/wiki/Mission_Statement

Page 34: Swapna Somasundaran swapna@cs.pitt

34

Crystal: Analyzing Predictive Opinions on the Web

Soo-min Kim and Eduard Hovy

Page 35: Swapna Somasundaran swapna@cs.pitt

35

Overview

• Crystal: Election prediction system– Messages on election prediction website– Predictive opinions – Automatically create annotated data– Feature generalization, Ngram features– Supervised learning

Page 36: Swapna Somasundaran swapna@cs.pitt

36

Outline

• Opinion types

• Task definition

• Data

• Results, Insights

Page 37: Swapna Somasundaran swapna@cs.pitt

37

Opinions

• Judgment Opinions• “I like it/ I dislike it”• Positive/Negative

• Predictive Opinions• “It is likely/ unlikely

to happen” • Belief about the

future• Likely/unlikely

Page 38: Swapna Somasundaran swapna@cs.pitt

38

Opinions

• Judgment Opinions

Sentiment Judgment, Evaluation, Feelings, Emotions

“This is a good camera”

“I hate this movie”

Page 39: Swapna Somasundaran swapna@cs.pitt

39

Opinions

• Predictive Opinions

Arguing (Wilson et. al, 2005, Somasundaran el al., 2007)– True (“Iran insists its nuclear program is for peaceful

purposes”)– will happen (“This will definitely enhance the sales”)– should be done (“The papers have every right to print them

and at this point the BBC has an obligation to print them.”)

Speculation (Wilson et al, 2005)– Uncertainty about what may/ may not happen

(“The president is likely to endorse the bill”)

Page 40: Swapna Somasundaran swapna@cs.pitt

40

Task

• Predictive Opinion – (Party, valence)

• Unit of prediction is message post on the discussion board

Page 41: Swapna Somasundaran swapna@cs.pitt

41

Page 42: Swapna Somasundaran swapna@cs.pitt

42

Data

• www.electionprediction.org

• Federal Election - 2004

• Calgary-east

• Edmonton-Beaumont

Page 43: Swapna Somasundaran swapna@cs.pitt

43

Data

• Gold standard: party logo used by author of the post– Positive examples– Negative examples?

Page 44: Swapna Somasundaran swapna@cs.pitt

44

Data

If you pick a party, all mentions of it => “likely to win”

If you pick a party, all mentions of

other parties => “not likely to win”

Page 45: Swapna Somasundaran swapna@cs.pitt

45

Page 46: Swapna Somasundaran swapna@cs.pitt

46

No tag LP=+1

Con= -1

No tag

Page 47: Swapna Somasundaran swapna@cs.pitt

47

Analyzing Prediction: Feature generalization

Similar to back-off

idea

Page 48: Swapna Somasundaran swapna@cs.pitt

48

Page 49: Swapna Somasundaran swapna@cs.pitt

49

Experiments

• Classify each sentence of the message

• Restore party names for “Party”

• Party with maximum valence is the party predicted to win by the message

Page 50: Swapna Somasundaran swapna@cs.pitt

50

Results

Baselines:• FRQ: most frequently mentioned party in the

message• MJR: most dominant predicted party• INC: current holder of the office• NGR: same as Crystal, only feature

generalization step is skipped• JDG: same as Crystal, but features are only

judgment opinion words

Page 51: Swapna Somasundaran swapna@cs.pitt

51

Results

•Crystal is the best performer at both the message and the riding level•Even with reduced features, crystal outperforms JDG system by ~ 4% points

Page 52: Swapna Somasundaran swapna@cs.pitt

52

Results: Insights

Page 53: Swapna Somasundaran swapna@cs.pitt

53

Results: Insights

Mutual Exclusivity

Mutual Exclusivity

Page 54: Swapna Somasundaran swapna@cs.pitt

54

Results: Insights

Sentiment

Page 55: Swapna Somasundaran swapna@cs.pitt

55

Results: Insights

desirability

Page 56: Swapna Somasundaran swapna@cs.pitt

56

Results: Insights

Modals Modals

Page 57: Swapna Somasundaran swapna@cs.pitt

57

Some Similar work

• Predicting Movie Sales from Blogger Sentiment, Mishne and Glance, (2006) AAAI-CAAW 2006

• Annotating Attributions and Private States, Wilson and Wiebe (2005). ACL Workshop 2005

• QA with Attitude: Exploiting Opinion Type Analysis for Improving Question Answering in On-line Discussions and the News , Somasundaran et al. ICWSM 2007.

Page 58: Swapna Somasundaran swapna@cs.pitt

58

Conclusion

• Explored predictive opinions

• Created automatically tagged election data

• Used feature generalization to train classifiers to predict election outcomes

Page 59: Swapna Somasundaran swapna@cs.pitt

59

Future work/Extensions

• Relation between judgment opinions and predictive opinions

• Other sentiment lexicons

• …?

Page 60: Swapna Somasundaran swapna@cs.pitt

60

Thank you!