Abstract: Calls to action on social media are known to be effective means of mobilization in social movements, and a frequent target of censorship. We investigate the possibility of their automatic detection and their potential for predicting real-world protest events, on historical data of Bolotnaya protests in Russia (2011–2013). We find that political calls to action can be annotated and detected with relatively high accuracy, and that in our sample their volume has a moderate positive correlation with rally attendance.
Introduction
While social media is a recognized engine for political mobilization, not all digital discourse contributes equally to offline action. "Calls to Action" (CTAs)—posts specifically urging users to join protests, disseminate information, or adopt symbolic gestures—represent the core of collective action. However, the clarity and urgency that make CTAs effective also make them prime targets for state-sponsored automated censorship. This research explores whether these critical posts can be reliably identified by machine learning and if their frequency serves as a predictive indicator for the scale of physical protest events.
Methodology and Taxonomy
To systematically study CTAs, the researchers first established a linguistic framework. They identified 15 distinct types of political CTAs, ranging from direct instructions for specific events to more subtle suggestions for collective action.
Using this taxonomy, three native Russian speakers annotated a subset of 1,000 posts from VKontakte (a major Russian social network) related to the Bolotnaya protests. The high Inter-Annotator Agreement (IAA) of 0.78 (Krippendorf’s alpha) confirmed that CTAs are a linguistically well-defined category.
For automated detection, the team tested several models, including traditional baselines like Logistic Regression (LR) and advanced deep learning architectures like RuBERT. The RuBERT model emerged as the top performer, achieving an F1-score of 0.78, though even simple LR models performed surprisingly well (0.71 F1), suggesting that CTAs possess a distinct vocabulary that is easily detectable even without state-of-the-art transformers.
Social Impact and Predictive Potential
By applying the best-performing model to a dataset of 91,000 historical posts, the researchers compared the volume of detected CTAs against Wikipedia-sourced attendance records for specific rallies. They found a moderate Pearson correlation of 0.4 between CTA volume and rally attendance, indicating that digital CTA frequency is a meaningful signal for "collective action potential."
The "Censorship Trap" and Chilling Effects
The study highlights a significant vulnerability: the very features that make CTAs effective for mobilization make them easy to suppress. The high detectability of these posts implies that censors do not require sophisticated AI to implement effective automated filters.
Furthermore, the researchers observed a "chilling effect" following the introduction of strict anti-protest regulations in mid-2012. Before these laws, engagement metrics like likes and reposts on CTAs spiked significantly before major events. Afterward, this digital activity stifled, demonstrating how legal pressure can decouple online mobilization signals from real-world participation.
Conclusion and Future Directions
The research confirms that political CTAs are a distinct, detectable, and impactful form of social media communication. While their volume correlates with physical mobilization, their high visibility poses a risk in restrictive political environments. Future work aims to develop tools for activists to test if their messaging is likely to trigger automated filters, potentially helping them find creative ways to bypass simple intent classifiers while still reaching their audience.
Summary updated