1. Quantitative Narrative Analysis of US Elections in
International News Media
Saatviga Sudhahar, Thomas Lansdall-Welfare, Ilias Flaounas,
Nello Cristianini

2. • Discourse about US Presidential Elections dominates the global
news system every 4 years
• Candidates take clear positions about a variety of issues, and many
social actors are expected to take sides in either endorsing or
opposing a candidate.
• The amount of news articles devoted to this topic is so large that no
exhaustive analysis can be attempted by conventional means
• We automate techniques from Quantitative Narrative Analysis
(QNA) for large scale narrative analysis of US elections in the global
media
Introduction

3. • QNA is aimed at identifying the actors and the actions that dominate
a story, as well as basic units of narration: triplets
Subject – Verb – Object (SVO) triplets
• This information captures a variety of relations that would be missed
by classical means
• identify Key actors and Key actions in elections
• automatically detect endorse/oppose relations between key
actors
• Generate a relational network whose topology depends on the
political relations between these players
• subject/object bias of actors in the political discourse
• subjects/objects of positive and negative statements

4. System Pipeline

5. Romney praised Paul Ryan. He recalled the excitement of the country in
electing Obama four years ago. Ryan criticized Obama for rejecting a deficit
reduction plan
Example
Romney praised Ryan. Romney recalled the excitement of the country in
electing Obama four years ago. Ryan criticized Obama for rejecting a deficit
reduction plan
Romney praise Ryan
Romney recall excitement
Ryan criticize Obama
After co-reference and anaphora resolution
Triplets extracted
Romney
Ryan
Obama
excitement
praise
criticizerecall

6. Endorse/Oppose relations
• Filter triplets that contain the key actors as subjects or objects; and
an endorse/oppose verb.
• Endorse verbs: appreciate, like, join, support
• Oppose verbs: criticize, hate, accuse, blame
• Each endorsement-relation between actors a, b is weighted by,
• , , denote the number of triplets between a, b that
support an endorse/oppose relation.

7. Network during
Primaries

8. Network after the
conventions

9. Filter reliable relations
• We consider a triplet reliable if it has been seen more than n times in
many articles.
• Higher threshold for n and retaining key actors in a network gives
more reliable information.
• We analyse party allegiance;
• that is the degree to which actors belong to one of two parties, in
the assumption that the election network is naturally organised
into two main communities.

10. The latest endorsement network after filtering
reliable relations

11. Network partitioning
• We use graph partitioning methods to analyse the allegiance of
actors to a party
• To perform its partitioning we first omit directionality by calculating
graph where is the adjacency matrix of the network
• We computed eigenvectors of and selected the eigenvector that
correspond to the highest eigenvalue.
• Elements of the eigenvector represent actors.
• We sort them by their magnitude and we obtain a sorted list of
actors.

13. Network from
2004 US election
Primaries

14. After the 2004
conventions

15. Spectrum of Actors showing party allegiance

17. Subject/Object Bias
• The Subject/Object bias of an actor reveals the role it plays in the
news narrative.

18. Positive/negative actors/campaign

20. Word Clouds of actions
Obama - Romney Romney - Obama

21. Election Watch:
http://electionwatch.enm.bris.ac.uk
719 US &International News Outlets.
processed1,48,104 articles.
extracted 4,80,952 triplets.

22. Evaluation
• We focus on the “high precision/low recall” setting
Validation 1:
• Compared system generated triplets with manually extracted ones
with a corpus containing Civil Rights movement in the Northern
Ireland
• Achieved 62% precision and 58% recall without applying filtering for
reliability.
• Probability of a triplet being incorrect - 0.38
• probability of error in triplets seen more than k times - 0.38k
• By only selecting triplets that are seen at least 3 times we achieve
5% error rate

23. Validation 2:
• We analysed by hand 75 triplets coming from the 2012 US Election
campaign, and checked how many were actually present in the
articles that were indicated by our pipeline as supporting them.
• Achieved 96% precision
• We have no estimation of recall, which we expect to be low.
Results on the past six election cycles on New York Times always
separated the two competing candidates along the eigenvector
spectrum.

24. Conclusion
• ElectionWatch presents key actors in U.S election news articles and
their role in political discourse
• The system is capable of detecting election-related articles, parsing
their content, solving co-reference and anaphora, identifying verbs
that denote support or opposition, identifying key actors, filtering
information that is statistically not reliable, and finally analysing the
properties of the resulting relational network.
• Future work will include
• making better use of the information coming from the parser,
which goes well beyond the simple SVO structure of sentences
• Develop more sophisticated methods for the analysis of large
and complex networks

25. Thank You
http://electionwatch.enm.bris.ac.uk
http://mediapatterns.enm.bris.ac.uk