Union restrictions on tribal groups have been a significant focus for two researchers who analyzed a multitude of articles published in Hungarian media between the collapse of 2021 and the spring of 2022. Martin Wendiggensen, a political scientist and doctoral candidate at Johns Hopkins University, collaborated with Benjamin Novak, a former reporter for The New York Times who is currently an undergraduate at the same university, to delve into the alignment of Hungarian online narratives with Russian disinformation outlets. Their investigation revealed a notable correlation between the two.
In September 2021, there was a surge in content promoting Russian interests in Hungary, preceding the actual invasion of Ukraine by Russian forces.
Presenting their findings at the recent LabsCon security event, Wendiggensen remarked, “We can only speculate on what drove the Hungarian media to closely mirror Russian propaganda thereafter.”
He noted a sharp escalation in articles covering three main themes from the fall of 2021 onwards, a trend that has persisted: criticism of weapons sales for prolonging conflicts, allegations of mistreatment of ethnic minorities by Ukraine, and the detrimental impact of European Union sanctions on Hungarian businesses.
Advancements in Machine Learning Analysis
Wendiggensen utilized a machine learning (ML) model to scrutinize the corpus of articles, while Novak conducted manual analysis. Their research is noteworthy for demonstrating the convergence of human and machine analysis, indicating the reliability of ML in detecting disinformation campaigns.
To gauge the country’s sentiment, Wendiggensen’s ML model tracked the prevalence of entire topics rather than individual words. Leveraging code blocks developed by ML expert Kohei Watanabe, the software autonomously extracted and categorized millions of articles into sections such as headlines, dates, and body text. By assigning multidimensional vectors to each of the 26 million terms collected, the software established semantic relationships based on vector positions and distances, enhancing the precision of word associations.
Wendiggensen elaborated on the model’s ability to identify semantic connections, citing examples like the close relationship between “sanctions,” “Brussels,” and “negative.” By mathematically computing vector distances, the model pinpointed the top three prevalent topics identified by Novak.
The ML model was further trained to assess polarity by contrasting opposing sentiments like “good” versus “bad,” using entire sentences to capture nuanced relationships. This approach enabled a holistic evaluation of viewpoints across multiple sentences, facilitating a comprehensive analysis of article sentiment within 15 minutes.
Wendiggensen emphasized the enduring dominance of the three main topics, underscoring the lack of media diversity in Hungary, which amplifies the dissemination of pro-Russian narratives. With media outlets predominantly aligned with the government, diverse viewpoints are constrained, fostering the prevalence of pro-Russian messaging.
Future Directions: Video Analysis and International Monitoring
In the next phase, Wendiggensen and Novak plan to analyze videos from Hungarian TV stations, expanding their dataset to include over 60 million terms transcribed from visual content. Additionally, they aim to extend their analysis to pan-European right-wing platforms, incorporating political affiliations to offer a comprehensive understanding of anti-European narratives.
Their ultimate goal is to create a dataset accessible for further research, enabling longitudinal studies on evolving communication patterns and the impact of economic conditions on media permissiveness. Wendiggensen envisions quantifying philosophical relationships to deepen insights into media dynamics over time.
