Understanding Russian information operations using unsupervised multilingual topic modeling

Abstract

What does this or that population think about a given issue? Which topics ‘go viral’ and why? How does disinformation spread? How do populations view issues in light of national ‘master narratives’? These are all questions which automated approaches to analyzing social media promise to help answer. We have adapted a technique for multilingual topic modeling to look at differences between what is discussed in Russian versus English. This kills several birds with one stone. We turn the data’s multilinguality from an impediment into a leverageable advantage. But most importantly, we play to unsupervised machine learning’s strengths: its ability to detect large-scale trends, anomalies, similarities and differences, in a highly general way. Applying this approach to different Twitter datasets, we were able to draw out several interesting and non-obvious insights about Russian cyberspace and how it differs from its English counterpart. We show how these insights reveal aspects of how master narratives are instantiated, and how sentiment plays out on a large scale, in Russian discourse relating to NATO.