Skip to content

Toggle service links

Persuasion across the Political Spectrum
Mr Martino Mensio

This event took place on 17th January 2024 at 11:30am (11:30 GMT)
Knowledge Media Institute, Berrill Building, The Open University, Milton Keynes, United Kingdom, MK7 6AA

In recent years, computational studies started analysing several aspects of News Media. The research community began developing models to extract and classify news articles in multiple dimensions: parallel news reports analysis, persuasion techniques detection, political leaning recognition and topic detection. While there is a lot of computational research on persuasion detection, we find that it has not been studied together with the other factors of parallel news reports, political leaning, and topics. This PhD analyses the use of language in parallel news articles to understand how persuasion changes across political leaning and topics. We use two methodologies in our work: (i) quantitative comparative analysis: we study one observed dimension across one or more partitioning variables; and (ii) classifier-based: we combine different variables as features for a classifier that needs to predict one additional variable. With these two methodologies, we analyse how persuasion changes across leanings and topics, and we see what is the effect of using persuasion and topics to predict the political leaning of news articles. Our work has three main contributions. (i) We improve the F1 of a political leaning classifier using the features of propaganda and topic. (ii) We discover that for specific topics, news sources use very different propaganda terms across the political spectrum, while still using the same techniques. (iii) We discover an imbalance in the standard datasets for propaganda detection, which raises questions about the generality of results in the literature. We believe that the results of this work can be useful both for computational researchers and users reading the news. For the first group, we emphasise the need to exploit the relationships discovered to improve models (using mixed features) and datasets (making them more representative of the news as a whole). For the second group, our work can be applied to build tools that would help the users compare how multiple articles use persuasion techniques differently on the same topic.

The webcast was open to 300 users
Creative Commons Licence KMi logo