More Related Content
Similar to Expression of Political Opinions in Press (20)
More from University Politehnica Bucharest (20)
Expression of Political Opinions in Press
- 1. Autor Conducător științific Universitatea Politehnica București Facultatea de Automatică și Calculatoare Catedra de Calculatoare Expression of Political Opinions in Press Costin-Gabriel CHIRU, Tudor Dimcica, Stere Caciandone [email protected]
- 2. Introduction • What? Application designed to analyze news articles from Romanian mass media and extract opinions about political entities • Why? study media polarization around important political events – elections • How? Using Machine Learning (ML) techniques for identifying and classifying opinions about the political entities • Results: reports and charts - used either for studying political polarization, or to identify partisan media 30.05.2017 CSCS21 2
- 3. Context • Over 50% of the world population connected to internet internet has one of the biggest exposures for information might replace the traditional media in the future • Digitization of the media people express their thoughts and feelings on the web producers of media content • Web users have immediate access to rich sources of information on any desired subject, both biased and unbiased (especially access to news information) • Past: newspapers were used to polarize large audiences • Present: migration of the audience into the online newspapers moving their content on the internet • Need and opportunity of tools for processing and extracting specific information to study how it affects and polarizes the audiences 30.05.2017 CSCS21 3
- 4. Analysis Steps • Extract information from given websites and save it into a database – using crawlers • Search for the entities of interest (input by the user) in the articles and create article-entity associations • Classify the opinions from the data in positive / negative using ML algorithms • Generate charts to show the computed polarities 30.05.2017 CSCS21 4
- 5. Application Architecture • Three main modules - data acquisition, data analysis and the graphical user interface (GUI) • Data acquisition – Extract only the political news from different newspapers’ website – Using a web crawler – Challenging because of the structural differences between different news websites – Some websites had a very chaotical structure couldn’t be crawled automatically eliminated – Kept for analysis: the title, the full content, the publishing date and the author. 30.05.2017 CSCS21 5
- 6. Data Analysis • Using 2 opinion classifiers from the scikit-learn library (Naive-Bayes & Support Vector Machines) • Challenges in choosing the granularity of the data (the level of the text analysis and classification) – Involved the detection of entities for which opinions were extracted • Search only specific entities that are chosen by the user • Could be automatized using a Romanian NER requires a lot of time should be done offline – Data normalization – transform the words in numbers • Bag of words + term frequency – Opinions detection - entity level • An article is labeled with opinions for each entity associated with it 30.05.2017 CSCS21 6
- 7. Graphical User Interface (GUI) • Main window - implements the functionality for classifying opinions and plotting visual statistics about the classified opinions • Article view - displays relevant data about an article and implements manual classification functionality 30.05.2017 CSCS21 7
- 9. Evaluation – Gold Standard • 50 random documents extracted from the dataset manually labeled by two students • Inter-rater agreement computed using the Cohen's kappa coefficient: 0.6019 (upper bound of the moderate agreement interval) 30.05.2017 CSCS21 9
- 10. Evaluation – Accuracy • Use gold standard to evaluate the performances of the algorithms used for classification: Naive- Bayes with multinomial distribution, Naive-Bayes with Bernoulli distribution, Support Vector Machines using Stochastic Gradient Descent • Algorithms chosen for good performance in categorizing texts 30.05.2017 CSCS21 10 Algorithms Students Student A Student B Average Naive-Bayes with multinomial distribution 66% 62% 64% Naive-Bayes with Bernoulli distribution 72% 64% 68% Support Vector Machines 80% 76% 78%
- 11. Case Study • 2014 Romanian Presidential Elections • Study focused on the two candidates that reached the second round: Victor Ponta and Klaus Iohannis • The source for opinion extraction was the news website Hotnews • Analysis before and after the elections 30.05.2017 CSCS21 11
- 12. Appearances in Press Before Announcing the Candidacy • Klaus Iohannis – member of the National Liberal Party – was the mayor of Sibiu from June 2000 until he was voted and invested as the president of Romania on December 2nd, 2014 30.05.2017 CSCS21 1 • Victor Ponta – jurist, former prosecutor, and a well- established political figure – former president of the Social Democratic Party – former Prime Minister of Romania – member of the Parliament since 2004 – Minister for Relations with Parliament between 2008 and 2009 – Prime Minister from May 2012
- 13. Opinions Extracted From Hotnews Articles • 2009 - 2013 30.05.2017 CSCS21 1 • Jan. 2014 – Sep. 2014
- 14. Appearances During Electoral Campaign and Elections • Appearances Oct. 2014 30.05.2017 CSCS21 14 • Appearances Nov. 2014 • Opinions Oct. 2014 • Opinions Nov. 2014
- 15. Results of Elections • First round took place on November, 2nd : Victor Ponta won 40.44%; Klaus Iohannis 2nd 30.37% • The second round took place on November 16th : Klaus Iohannis won with 54.43%, although the opinion polls made with traditional methods presented Victor Ponta as the winner of the election by a high margin • Klaus Iohannis became the politician with the most followers on Facebook in the entire Europe and he enjoyed a high degree of positive opinions in the online media. 30.05.2017 CSCS21 15
- 16. Appearances After Elections 30.05.2017 CSCS21 16 • Post-Election Appearances • Post-Election Opinions
- 17. Conclusions (1) • Tool for sentiment analysis / opinion mining for Romanian online media • The user may add / remove entities from the analysis • The user can manually label the opinion of articles towards entities or he/she can use the automatic labeling using ML algorithms • The generated charts offered interesting findings that support the results of the analyzed presidential elections • Using this tool complementary with traditional polls can create a very good picture of how opinions are polarized around political entities 30.05.2017 CSCS21 17
- 18. Conclusions (2) • Main limitation – The use of only this tool isn't enough, as the gap between the popularity of the two candidates was much higher than the difference of votes between them in the second round • Biggest challenge - find a suitable way to analyze the available data • Biggest advantage - capability of processing large quantities of data and extracting unbiased opinions • Possible improvements: crawlers for other newspapers, publish it online, export the charts in different formats, use different weights for articles based on popularity / visibility of the source or the author, extend the range of analyzed sentiments 30.05.2017 CSCS21 18
- 19. Questions 30.05.2017 CSCS21 19 Thank you very much!