The Taming of News Flows. A Deep Learning Approach to Mapping the Discourse of Foreign News in Romania and Hungary


With journalism and news organizations looking to resist increasing economic pressure brought about by digital media giants’ disruptive ecosystems, foreign news reporting relies ever more on global agencies. However, selection processes applied to news flows may point to particular national-level, culture-specific or organizational tendencies to filter and adapt news to local contexts. What are the dominant news values and topics preferred when foreign news are reported?

Romania and Hungary, although neighboring countries in Central and Eastern Europe with shared history, have some distinctive features with respect to their position and positioning in the European Union, political leadership, media system ownership structures, ties with neighboring countries and public perceptions of regional and global issues. Investigating foreign news reported by the Romanian and Hungarian online media provides insight into different maps of meaning that news outlets, journalists and their publics use to understand the world.

This research investigates international news coverage over a large dataset (N=110.075) of articles published by three leading online news outlets in Romania (N1=41.363) and Hungary (N2=68.712) over the course of three years – 2016, 2017 and 2018. Grounded in the news flows/news factors approach proposed by Galtung and Ruge (1965) and building on research carried out on foreign news by researchers such as Wu(2000), Hanusch and Obijiofor (2008) and Segev (2015), this approach advances state of the art research by leveraging type-based techniques of deep learning text classification to provide an overview of news values, topics and represented countries/world areas in the Romanian and Hungarian media. Bednarek & Caple (2014) define news values and their key linguistic devices for use in content analyses, but recent developments in machine learning allow for attempts to automate the annotation of news values and topics (according to the IPTC Media Topics top level categories) in research over large corpora. Furthermore, our supervised learning approach makes use of country names and news agency sources derived from lexicon-based automated coding, besides the publications, titles, full text, authors, dates and keywords to enhance the performance of the annotation. This research attempts to bridge between approaches recently developed in computer science with research into news flows, news values and the domestication of foreign news developed by Gurevitch et al. (1991). The analysis of topics, news values and references (countries, persons, entities, sources) allows identification of domestication strategies specific to either of the two countries’ media.