Identifying Ideological Perspectives of Web Videos using Patterns Emerging from Folksonomies Wei-Hao Lin and Alex Hauptmann Language Technologies Institute School of Computer Science Carnegie Mellon University Wikis, Blogs, Bookmarking Tools - Mining the Web 2.0 Workshop @ ECML 2008
YouTube videos about abortion issue
YouTube videos about abortion issue
YouTube videos about abortion issue
YouTube videos about abortion issue “Pro-choice” view
YouTube videos about abortion issue “Pro-choice” view
“Pro-life” view
YouTube videos about abortion issue “Pro-choice” view
pro, choice, feminism, abortion, women, rights, truth, Bush
“Pro-life” view
YouTube videos about abortion issue “Pro-choice” view
“Pro-life” view
pro, choice, feminism, abortion, women, rights, truth, Bush
prayer, pro-life, God
Echo Chamber Percentage of Internet users in the United States who seek news sources that challenge their views:
Echo Chamber Percentage of Internet users in the United States who seek news sources that challenge their views:
20%
Too many videos, too little time
Automatic detection of biased web videos
Automatic detection of biased web videos Warning: Strongly pro-life bias!
Automatic detection of biased web videos Warning: Strongly pro-life bias! Want to watch pro-choice videos?
Goal
•
Automatic identify a web video’s ideological viewpoint
• • •
Justify statistical models’ decision Raise awareness of individual video makers’ biases Facilitate mutual understanding between people holding different beliefs
Outline
• • • •
Goal: Identify web videos’ ideological perspectives Joint Topic and Perspective Model
• •
Emphatic patterns in user-generated tags Model specification
Evaluation Conclusions
Emphatic patterns in tag multinomial ß
• • •
Two videos on different events: very different ß Two videos from same views: similar ß Two videos from differing views: different ß
v.s.
Emphatic patterns in tag multinomial ß
• • •
Two videos on different events: very different ß Two videos from same views: similar ß Two videos from differing views: different ß
v.s.
Emphatic patterns in tag multinomial ß
• • •
Two videos from same views: similar ß Two videos on different events: very different ß Two videos from differing views: different ß
v.s.
Encode emphatic patterns into ß structure
Encode emphatic patterns into ß structure
Joint Topic and Perspective Model (jTP) µτ τ π
Pd
Wd,n
Στ
βv Nd D
Document view Pd ∼
V
µφ
φv V
Σφ
Bernoulli(π), d = 1, . . . , D
word Wd,n |Pd = v ∼ Multinomial(βv ), n = 1, . . . , Nd βvw =
topical weight τ ∼
ideological weight φv ∼
exp(τ w ×φw v ) P ! ! ,v w w ×φv ) w! exp(τ
= 1, . . . , V
N(µτ , Στ ) N(µφ , Σφ ).
Joint Topic and Perspective Model (jTP) µτ τ π
Pd
Wd,n
Στ
βv Nd D
Document view Pd ∼
V
µφ
φv V
Σφ
Bernoulli(π), d = 1, . . . , D
word Wd,n |Pd = v ∼ Multinomial(βv ), n = 1, . . . , Nd βvw =
topical weight τ ∼
ideological weight φv ∼
exp(τ w ×φw v ) P ! ! ,v w w ×φv ) w! exp(τ
= 1, . . . , V
N(µτ , Στ ) N(µφ , Σφ ).
Joint Topic and Perspective Model (jTP) µτ τ π
Pd
Wd,n
Στ
βv Nd D
Document view Pd ∼
V
µφ
φv V
Σφ
Bernoulli(π), d = 1, . . . , D
word Wd,n |Pd = v ∼ Multinomial(βv ), n = 1, . . . , Nd βvw =
topical weight τ ∼
ideological weight φv ∼
exp(τ w ×φw v ) P ! ! ,v w w ×φv ) w! exp(τ
= 1, . . . , V
N(µτ , Στ ) N(µφ , Σφ ).
Joint Topic and Perspective Model (jTP) µτ τ π
Pd
Wd,n
Στ
βv Nd D
Document view Pd ∼
V
µφ
φv V
Σφ
Bernoulli(π), d = 1, . . . , D
word Wd,n |Pd = v ∼ Multinomial(βv ), n = 1, . . . , Nd βvw =
topical weight τ ∼
ideological weight φv ∼
exp(τ w ×φw v ) P ! ! ,v w w ×φv ) w! exp(τ
= 1, . . . , V
N(µτ , Στ ) N(µφ , Σφ ).
Joint Topic and Perspective Model (jTP) µτ τ π
Pd
Wd,n
Στ
βv Nd D
Document view Pd ∼
V
µφ
φv V
Σφ
Bernoulli(π), d = 1, . . . , D
word Wd,n |Pd = v ∼ Multinomial(βv ), n = 1, . . . , Nd βvw =
topical weight τ ∼
ideological weight φv ∼
exp(τ w ×φw v ) P ! ! ,v w w ×φv ) w! exp(τ
= 1, . . . , V
N(µτ , Στ ) N(µφ , Σφ ).
Joint Topic and Perspective Model (jTP) µτ τ π
Pd
Wd,n
Στ
βv Nd D
Document view Pd ∼
V
µφ
φv V
Σφ
Bernoulli(π), d = 1, . . . , D
word Wd,n |Pd = v ∼ Multinomial(βv ), n = 1, . . . , Nd βvw =
topical weight τ ∼
ideological weight φv ∼
exp(τ w ×φw v ) P ! ! ,v w w ×φv ) w! exp(τ
= 1, . . . , V
N(µτ , Στ ) N(µφ , Σφ ).
Outline
• • •
•
Goal: Identify web videos’ ideological perspectives Joint Topic and Perspective Model Evaluation
• • •
YouTube videos on 8 issues Uncovered topical and ideological weights Classifying a video’s viewpoint
Conclusions
Experimental Data
• •
Download top 1K YouTube videos and associated tags 8 political and social issues 1.
Abortion
2.
Democratic party primary election in 2008
3.
Gay rights
4.
Global warming
5.
Illegal immigrants to the United States
6.
Iraq War
7.
Israeli-Palestinian conflict
8.
United States politics
Pro-life vs. Pro-choice
Global warming supporters vs. skeptics
Good accuracy on identifying viewpoints
Conclusions
• • •
•
New challenge: identifying ideological perspectives Emphatic patterns in user-generated tags Joint topic and perspective model
•
Automatic discovery of topical and ideological weights of tags
•
Achieve good accuracy on identifying viewpoints
Emphatic patterns prevail in multiple modalities: e.g., words and visual content