Using Interactions in the Quantification of Media Bias

DP
by Diogo Pacheco0 views·Updated Oct 9, 2014
Powered by AI

Need to present on this topic?

Create your own polished presentation in minutes. Just describe what you need.

Free to try, no credit card required

Slide outline

  1. T'=[ ] — 13
  2. D1=[ ] — Example:
  3. D1=[ ] — Example:
  4. T=[ ] — Example:
  5. M1=[ ] — 11
  6. D2=[ ] — .13
  7. M2=[ ] — 15
  8. T=[ ] — 10
  9. =40/30 = 1.33
  10. D1=[ ] — .1
  11. D2=[ ] — .13

Transcript

Friday, October 10, 2014

$Free

Vol IX, No. 2

Outline

Using Interaction in the Quantification of Media Bias

Experimental Results

Conclusions

Future Works

Introduction

Related Work

Methodology

Quantifying Coverage

General Coverage

Bias Classification

Data Collection

Diogo F. Pacheco*, Dillon Rose*, Fernando B. Lima-Neto**, Ronaldo Menezes*

* BioComplex Laboratory, Computer Science, Florida Institute of Technology, USA

** Escola Politécnica, University of Pernambuco, Brazil

dpacheco2013@my.fit.edu, drose2010@my.fit.edu, fbln@ecomp.poli.br, rmenezes@cs.fit.edu

Presented in 2014 ASE BIGDATA/SOCIALCOM/CYBERSECURITY Conference, Stanford University.

Introduction

Media outlets portray themselves as “neutral” or “nonpartisan”

Does bias really exist? If yes, can it be quantified?

Interpretation for coverage of media outlets evaluate media bias

Evaluate without a reference point – a ground truth

The evaluation is so reliable as the ground truth

“True reports”, “false reports”, and “lack of reports” compose the media coverage

Related Work

Introduction

Introduction

The importance of our results are twofold:

  • Citizens can use such metrics to evaluate the quality of the news they read
  • Newspapers could use our approach to gauge themselves and adjust their coverage

All these studies provide absolute results and are not based on a “truth”

This can be a concern because favoritism refers to doing something abnormal and not just a count

If a party works more, this party should get more attention and that would be considered normal if this imbalance is proportional

Selected 10 newspapers of the largest in the USA

Politics (US Senate) as an example

Nature of collaboration within the Senate give us a “truth” target

This approach may be controversial because of bills’ importance, but this popularity may be a byproduct of media coverage

Liberals say the media is conservative [5] and conservatives say the media is liberal [1]

D’Alessio and Allen [3] investigated 3 kinds of bias in newspapers, magazines, and television. Only television had slight bias.

Druckman and Parkin [6] showed the effect of a newspaper tone towards a candidate can impact voters’ decision

Groceclose and Mulyo [8] proposed a measure of media bias based on a count of mentions to specific think tanks. The starting point is the ADA metric (Americans for Democratic Actions)

their collaboration should also be reported

If two senators collaborate in the Senate,

[5] E. E. Dennis. Liberal media? not hardly. St. Petersburg Times, Nov. 1992. Page 1D.

[1] L. B. Bozell. How can liberals deny the media's slanted coverage? Philadelphia Inquirer, Oct. 1992. Page 9.

[3] D. D'Alessio and M. Allen. Media bias in presidential elections: A meta-analysis. Journal of Communication, pages 133{156, Aug. 2000.

[6] J. N. Druckman and M. Parkin. The impact of media bias: How editorial slant affects voters. Journal of Politics, 67(4):1030{1049, 2005.

[8] T. Groseclose and J. Milyo. A measure of media bias. The Quarterly Journal of Economics, CXX(4), Nov. 2005.

treat equals equally and unequals unequally in proportion to the relevant difference

Aristotle

= Legitimacy + Fabrication + Negligence

Quantifying Coverage

General Coverage

General Coverage

Medias and the ground truth may be represented in “different magnitudes”, so it is necessary scale the truth to make it comparable with Medias

Adjustment factor based on maximum value

Then we calculate the matrix of relative differences. It represents the number of times a relationship is over/under covered

With matrix of differences and a threshold we classify each relationship

The media coverage will be represented by the amount of legitimate, fabricated, and neglected reports.

Coverage allow us to investigate favoritism of any subject with social relationship based on differences between measured and expected

We decided to evaluate our proposed method from the angle of “collaborativeness” – social network

Compare the ground truth social network with media social networks. Every relationship is measured as:

We proposed two approaches to measure these three criteria (coverage composition): naive (binary) and general (weighted)

Threshold 5%:

Leg. = 0 -> 20%

Fab. = .13, 6.5 -> 40%

Neg. = -.25, -20 -> 40%

Threshold 15%:

Leg. = .13, 0 -> 40%

Fab. = 6.5 -> 20%

Neg. = -.25, -20 -> 40%

Threshold 5%:

Leg. = 0 -> 20%

Fab. = .1, 2 -> 40%

Neg. = -.1, -2 -> 40%

Threshold 15%:

Leg. = .1, -.1, 0 -> 60%

Fab. = 2 -> 20%

Neg. = -2 -> 20%

Data Collection

Data Collection

Bias Classification

Evaluate media coverage through collaboration

We used the New York Times Congress API to gather cosponsorship data from 113th US Senate - the ground truth

It may be useful, sometimes, to have a finer granularity of the coverage

In the case of politics, another perspective can be extracted from the matrix of differences

The coverage could be:

  • Favoring a party, an ideology – high level bias
  • Skewed towards particular individuals – low level bias

We used google’s search to measure the strength of social relation between senators as the number of common citation for every newspaper

We also collected data from senate.gov website

There is no correlation between circulation and citations

Some newspaper are more “social” than others

Democrats:

(.1 - .1) / 2 = 0

Republicans:

(-2 + 0 + 2) / 3 = 0

Democrats:

(.13 - .25) / 2 = -.06

Republicans:

(-20 + 0 + 6.5) / 3 = -4.5

Experimental Results

Experimental Results

Experimental Results

Coverage Quantification

Bias Classification - Wrong Evaluation

Bias Classification

It is well known that the Senate has been evenly represented by democrats and republicans for several years

  • Assumption 1: they work in the same proportion they are represented
  • Assumption 2: newspapers are unbiased

There is no best newspaper because readers may use different criteria to express what is important

Conclusions

Future Works

Investigate more types of media (radio, television)

Evaluate if the senate.gov can be used as a ground truth

Try to use NLP to generate more accurate citation networks

Investigate collaboration not only in 113th Senate since the citation network data are not time constrained

We proposed a metric called coverage inspired by three criteria: legitimacy, fabrication, and negligence

We used a collaboration network from the US Senate to compare with citation networks from 10 major newspapers

Coverage is not correlated to circulation

There is no best newspaper since readers may prefer one criterion rather than the other

The Dallas Morning News is the only newspaper favoring republicans

The NY Times and the senate.gov were classified as uncertain bias

1

T'=[ ]

13

27

20

40

7

2

Example:

.1

-.1

-2

0

2

D1=[ ]

3

Example:

John McCain | Mark Begich | Tom Udall | Mark Kirk | . . .

John McCain

Mark Begich

Tom Udall

Mark Kirk

.

.

.

4

Example:

.1

-.1

-2

0

2

D1=[ ]

D2=[ ]

.13

-.25

-20

0

6.5

5

Example:

T=[ ]

10

20

15

30

5

6

fair

work

bills

democrats

republicans

media

trust

citation

collaboration

cosponsorship

7

M1=[ ]

11

18

5

30

15

8

Legitimate when appears in media proportionally to the ground truth – “true reports”

Fabricated when appears in media and does not exist, or appears much more than expected – “false reports”

Neglected when does not appear in media and exists, or appears much less than expected – “lack of reports”

9

D2=[ ]

.13

-.25

-20

0

6.5

10

15

20

0

40

50

M2=[ ]

11

T=[ ]

10

20

15

30

5

12

=40/30 = 1.33

13

.1

-.1

-2

0

2

D1=[ ]

14

D2=[ ]

.13

-.25

-20

0

6.5

15

Individual Level (Fabrication):

.1 < 2 -> .13 < 6.5 ->

16

Example:

John McCain | Mark Begich | Tom Udall | Mark Kirk | . . .

John McCain

Mark Begich

Tom Udall

Mark Kirk

.

.

.