Animation showing movements of key actors in presidential campaign, as seen through automated content analysis of 148574 news articles. SIze reflects number of mentions, vertical axis reflects subject/object bias, color reflects positive/negative bias.
Quantitative Narrative Analysis of US Elections in International News Media
Quantitative Narrative Analysis of US Elections inInternational News MediaSaatviga Sudhahar, Thomas Lansdall-Welfare, Ilias Flaounas,Nello Cristianini
• Discourse about US Presidential Elections dominates the globalnews system every 4 years• Candidates take clear positions about a variety of issues, and manysocial actors are expected to take sides in either endorsing oropposing a candidate.• The amount of news articles devoted to this topic is so large that noexhaustive 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 globalmediaIntroduction
• QNA is aimed at identifying the actors and the actions that dominatea story, as well as basic units of narration: tripletsSubject – Verb – Object (SVO) triplets• This information captures a variety of relations that would be missedby classical means• identify Key actors and Key actions in elections• automatically detect endorse/oppose relations between keyactors• Generate a relational network whose topology depends on thepolitical relations between these players• subject/object bias of actors in the political discourse• subjects/objects of positive and negative statements
Romney praised Paul Ryan. He recalled the excitement of the country inelecting Obama four years ago. Ryan criticized Obama for rejecting a deficitreduction planExampleRomney praised Ryan. Romney recalled the excitement of the country inelecting Obama four years ago. Ryan criticized Obama for rejecting a deficitreduction planRomney praise RyanRomney recall excitementRyan criticize ObamaAfter co-reference and anaphora resolutionTriplets extractedRomneyRyanObamaexcitementpraisecriticizerecall
Endorse/Oppose relations• Filter triplets that contain the key actors as subjects or objects; andan 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 thatsupport an endorse/oppose relation.
Filter reliable relations• We consider a triplet reliable if it has been seen more than n times inmany articles.• Higher threshold for n and retaining key actors in a network givesmore reliable information.• We analyse party allegiance;• that is the degree to which actors belong to one of two parties, inthe assumption that the election network is naturally organisedinto two main communities.
The latest endorsement network after filteringreliable relations
Network partitioning• We use graph partitioning methods to analyse the allegiance ofactors to a party• To perform its partitioning we ﬁrst omit directionality by calculatinggraph where is the adjacency matrix of the network• We computed eigenvectors of and selected the eigenvector thatcorrespond to the highest eigenvalue.• Elements of the eigenvector represent actors.• We sort them by their magnitude and we obtain a sorted list ofactors.
Word Clouds of actionsObama - Romney Romney - Obama
Election Watch:http://electionwatch.enm.bris.ac.uk719 US &International News Outlets.processed1,48,104 articles.extracted 4,80,952 triplets.
Evaluation• We focus on the “high precision/low recall” settingValidation 1:• Compared system generated triplets with manually extracted oneswith a corpus containing Civil Rights movement in the NorthernIreland• Achieved 62% precision and 58% recall without applying filtering forreliability.• 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 achieve5% error rate
Validation 2:• We analysed by hand 75 triplets coming from the 2012 US Electioncampaign, and checked how many were actually present in thearticles 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 alwaysseparated the two competing candidates along the eigenvectorspectrum.
Conclusion• ElectionWatch presents key actors in U.S election news articles andtheir role in political discourse• The system is capable of detecting election-related articles, parsingtheir content, solving co-reference and anaphora, identifying verbsthat denote support or opposition, identifying key actors, filteringinformation that is statistically not reliable, and finally analysing theproperties 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 largeand complex networks