slides - The Stanford NLP

advertisement

Who Needs Polls?

Gauging Public Opinion from Twitter Data

David Cummings

Haruki Oh

Ningxuan (Jason) Wang

From Tweets to Poll Numbers

• Motivation: People spend millions of dollars on polling every year: politics, economy, entertainment

• Millions of posts on Twitter every day

• Can we model public opinion using tweets?

• Data: 476 million tweets from June to December

2009, courtesy of Jure Lescovec

• Public polls from The Gallup Organization (presidential approval, economic confidence) and Rasmussen

Reports (generic Congressional ballot)

• Goal: high correlation with public opinion polls

• All correlation figures for 6-day smoothing window

Approach 1: Volume

• The simplest metric: percentage of tweets that mention a given topic in a certain time window

• Moderate negative correlation (-36.3%, -35.7%) for economy and Congressional ballot: mention things you want to complain about more often

• Higher correlation (52.4%) for Obama

Presidential Approval: Volume Metric

35

30

25

20

15

10

5

0

06

/1

4

06

/2

2

06

/3

0

07

/0

8

07

/1

6

07

/2

4

08

/0

1

08

/0

9

08

/1

7

08

/2

5

09

/0

2

09

/1

0

09

/1

8

09

/2

6

10

/0

4

10

/1

2

10

/2

0

10

/2

8

11

/0

5

11

/1

3

11

/2

1

11

/2

9

12

/1

5

0.9

0.8

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0 approval volume

Approach 2: Generic Sentiment

• Can we distinguish between positive and negative sentiment of tweets?

• University of Pennsylvania OpinionFinder subjective polarity lexicon

• “conceited” strong negative -10

• “ironic” weak negative -5

• “trendy” weak positive +5

• “illuminating” strong positive +10

• Sum word scores for a tweet to classify it as positive, negative, or neutral; then subtract negative counts from positive counts and normalize over window

Approach 2: Generic Sentiment

• Good results on economic confidence: 60.4% correlation, 70.1% correlation on 15-day window

• Poor performance on presidential approval and

Congressional ballot: -24.5% and 21.5% correlation respectively

• Sentiment about politics expressed differently?

Economic Confidence: Generic Sentiment

-15

-20

-25

-30

-35

-40

06

/1

4

06

/2

2

06

/3

0

07

/0

8

07

/1

6

07

/2

4

08

/0

1

08

/0

9

08

/1

7

08

/2

5

09

/0

2

09

/1

0

09

/1

8

09

/2

6

10

/0

4

10

/1

2

10

/2

0

10

/2

8

11

/0

5

11

/1

3

11

/2

1

11

/2

9

12

/0

7

12

/1

5

12

/2

3

0.4

0.35

0.3

0.25

0.2

0.15

conf idence sentiment

Approach 3: LM-based Classification

• Train three language models (positive, negative, and neutral) on hand-classified data

• Classify each tweet according to the language model that affords it the highest probability

• Applied for the case of Obama: manually classified

3,633 tweets

• “can we all talk about how awesome Obama is?”

• “that Obama sticker on your car might as well say ‘Yes

I’m stupid’ #tcot #iamthemob #teaparty #glennbeck”

• Then we tested the language models: best performer was a linearly interpolated bigram model

Approach 3: LM-based Classification

• Much-improved results on presidential approval:

49.4% correlation

• Throwing out retweets and duplicate tweets helps a little more: 55.9% correlation

• Finally, combining both volume and LM-based sentiment gives best results: 63.3% correlation, or

69.6% correlation on a 15-day window

Presidential Approval: LM-based Classification

35

30

25

20

15

10

5

0

06

/1

4

06

/2

3

07

/0

2

07

/1

1

07

/2

0

07

/2

9

08

/0

7

08

/1

6

08

/2

5

09

/0

3

09

/1

2

09

/2

1

09

/3

0

10

/0

9

10

/1

8

10

/2

7

11

/0

5

11

/1

4

11

/2

3

12

/0

2

12

/1

1

12

/2

0

12

/2

9

0.1

0.05

0

-0.05

-0.1

-0.15

-0.2

-0.25

-0.3

approval

LIBigram

LIBigram--

Download