Analyzing Audience Responses to Ukrainian Memetic Warfare Tactics

Table of Links

Introduction

Background and Related Work

Methods

Results

Discussion and References

4 RESULTS

We now analyze this annotated dataset to answer our research questions. Section 4.1 outlines which narratives, emotions, and actors are more prevalent in each of the three accounts. Section 4.2 examines which factors are more important for the virality of this content. Finally, Section 4.3 shows how the audience breaks down across countries.

4.1 Content analysis

Figure 1 outlines which elements are more prevalent in each of the considered accounts. Unsurprisingly, @DefenceU focuses on the heroic narrative by depicting the Ukrainian military as powerful and morally good. Instead, the independent account @uamemesforces focuses on the fool narrative, where the other side is depicted as weak and incompetent. At the same time, the account employs a large amount of hero narrative, partially due to the memes of Ukrainian farmers resisting Russian soldiers.

The actors mentioned by each account also differ. @DefenceU puts the spotlight on the Ukrainian military but also dedicates some attention to Ukrainian civilians. Interestingly, Putin is mentioned by @uamemesforces much more than @DefenceU, which almost never mentions him. Similarly, actors associated with the West are mentioned more by @uamemesforces, while @DefenceU tends to ignore them. @Ukraine focuses the most on Ukrainian civilians and Ukraine as a nation, possibly working towards the goal of nation-building, or to keep the global public opinion focused on the main victims of the conflict. The three accounts are also distinct in their emotional appeal: @Ukraine communicates largely outrage and pride, while @uamemesforces focuses on humor, and @DefenceU on pride alone. Concerning possible intents (not plotted), @Ukraine posts by far the most calls to action (31%), with many calls to support the war effort, while @DefenceU often posts purely informational material (e.g., war bulletins).

Next, we examine the role of the main actors found in the narratives. Figure 2 presents heat maps of the intersection of actors and narratives for each account. Across all three accounts, the dichotomy between the two sides of the conflict is stark: the Russian side is portrayed as either a fool or a villain (negative moral quality), while the Ukrainian side as either a hero or a victim (positive moral quality); the West spans both portrayals. The emphasis by @Ukraine is on Ukrainian civilians and Ukraine as a whole, portrayed in a victim role, whereas less emphasis is put on the actors from the Russian side. Instead, @uamemesforces explores both sides at length, including various Western actors. The most common combination of actor/narrative is portraying the Russian military as a fool (negative moral quality and powerless), and a similar treatment is reserved for Putin. This powerless portrayal is much more common than the powerful one (in the role of a villain). Unlike for @Ukraine, @uamemesforces content rarely portrays Ukraine as a victim, even when talking about civilians, and instead attributes them heroic characteristics: a recurring example of this combination are references to a story of Ukrainian farmers who towed away abandoned Russian tanks. Actors from

the West are mostly portrayed as fools or heroes—mostly depending on whether the particular actor is perceived as supporting Ukraine—but in a few cases, they are even portrayed as villains, when they are perceived to be complicit with the enemy. The narrative of @DefenceU consists of several well-defined combinations: Ukrainian military as heroes, Russian military as fools, and Ukrainian civilians as victims. It is noteworthy that the Russian military is often portrayed as powerless (fool), whereas Russia as a country is always portrayed as a villain (powerful), and Putin himself is rarely referred to personally (contrary to the practice of @uamemesforces). Russian civilians are only briefly mentioned by @uamemesforces (omitted from the plots, 20 tweets in total), often as “conniving” with their government. Thus, we find distinctive narrative styles in the three accounts: @Ukraine focuses on its victimhood, @uamemesforces emphasizes the weakness (foolishness) of the Russian side and the heroism of Ukrainians, and @DefenceU propagates a heroic, powerful view of its own military while portraying their enemy in a foolish, weak perspective.

Finally, Figure 3 shows the emotional appeal of the posts in each account. Since a tweet labeled with a single emotion could refer to several actors, in our analysis we associate such emotion with all the actors present. The emotional focus is very well defined for @DefenceU, where pride is present in an overwhelming majority of the posts. A combination of humor dominates the messaging style of @uamemesforces. Interestingly, Putin is singled out as an object of ridicule only by @uamemesforces.

Examples. Let us now focus on the narratives, given their explanatory power that we explore in Section 4.2. Here we give a few examples to crystallize the imagery associated with each narrative. These examples are taken from the top 10 most popular tweets per narrative.

Figure 4 illustrates examples of ‘hero’ narrative. This narrative typically depicts the heroism of Ukrainian military efforts, emphasizing their benevolent moral quality and their power to defeat the enemy. In fact, it is often associated with mentions of Ukraine as an actor. A typical example is the leftmost in Figure 4. This meme was posted by @uamemesforces on 28 March 2023; in the same days, Ukrainian officials were claiming victories against the Russian military forces, such as the city of Irpin being recaptured. This meme by @uamemesforces supports the same narrative, but uses typical memetic language: the moment at the 2022 Oscars when Will Smith slapped Chris Rock, whose videos gained tens of millions of views, is edited to show Zelensky fiercely slapping Putin. This tweet obtained more than 120k likes. Other content, typically by @Ukraine, promotes information on ways to support Ukraine, or celebrates the valor of Ukrainian

civilians, such as their victory in the Eurovision Song Contest, and connects it to the “brave freedom defenders” on the frontline (middle example in the figure). @DefenceU focuses on the military: not-so-veiled military threats (e.g., a video of a military toy floating by the Kerch bridge, not shown) and references to an online social-media organization dedicated to countering Russian online propaganda (the “North Atlantic Fellas Organization #NAFO”, rightmost image).

Figure 5 shows examples of ‘victim’ narrative. The most popular ones are from @Ukraine: for instance, the news of the world’s largest plane, the Ukrainian Antonov AN-225, having been destroyed by Russians is used to illustrate the sufferings of the war (leftmost). Another example is shown in the middle, where @Ukraine uses a photo of civilians whose home was destroyed by Russian bombings in the city of Chuhuiv, near Kharkiv, to ask the global public opinion support for a specific economic sanction towards Russia—the closing of seaports. The last example is from @DefenceU, who uses a poster-like design to promote their view that Russia’s actions in Ukraine fall under the definition of genocide—something that the United Nations is currently investigating and that at the time of writing is inconclusive.[10]

Figure 6 displays examples of ‘villain’ narrative, focusing on the immorality of the enemy and its strength. They often juxtapose Russia with Ukraine and emphasize the civil casualties incurred by Ukraine. Negative historical associations (e.g., with Stalin or Hitler) are often employed (leftmost example). Extreme evilness is attributed to the Russian army: in the middle example, Russian pilots are depicted by a cartoon as gloating when killing Ukrainian babies, a typical atrocity tale employed in propaganda [12]. Even Russian civilians are shown as evil: in the rightmost example, they happily play with toys stolen from Ukrainian children.

Figure 7 presents examples of ‘fool’ narrative, which still focuses on the immorality of the enemy, but depicts it as weak and powerless. Many examples focus on the failed ambitions of Russia and depict their soldiers as incompetent and ineffective, as typical in psychological warfare.

Tweets containing ‘other’ narratives are often associated with humorous posts and mention actors other than Ukraine, bearing an ambivalent attitude. Figure 8 shows several examples, which provide a commentary on the international response to the war, the sanctions, and the role of Hungary in the conflict. While an overarching narrative can be inferred, it does not clearly fit within the narrative framework we employ.

4.2 Modeling popularity

We now move to modeling the popularity of the content shared by these accounts. We use a linear regression with the content variables outlined above as independent variables. Categorical variables are represented via one-hot encoding with “none” as the reference case. The target variable—the number of retweets—is transformed via log base 2, such that a coefficient of 1 indicates the number of retweets doubles, while a -1 that it halves.

First, we investigate which types of features are more effective in predicting the success of a tweet. Table 2 shows the performance of models with different subsets of explanatory variables. All the models include the ‘baseline’ variables: the identity of the author, the number of their followers, and the number of days from the invasion. The other models use the variables grouped as visual, actor, emotion, intent, and narrative. We perform model selection based on the Bayesian Information Criterion (BIC), a metric that captures both unexplained variation in the dependent variable and the number of explanatory variables. The best-performing model uses visual and narrative variables, and achieves −43.9 ΔBIC compared to the baseline, with an Adjusted 𝑅2 of 0.622. Remarkably, the Narrative model alone (with only 5 features) performs nearly as well as the Visual model (state of the art with 16 features) in terms of Adjusted 𝑅2 , and even better in terms of ΔBIC. This result testifies to the importance of narrative features in determining content popularity.

The coefficients for the best model and their 99% confidence intervals are shown in Figure 9. The narrative-related variables have some of the largest positive coefficients, with the exception of screenshot types of images and long-scale images. The fact that all narrative coefficients are positive implies that the very existence of a narrative is associated with greater engagement in terms of retweets. Out of these narratives, the victim one is the most resonant: its presence is associated with an increase of 109% in the number of retweets. The other benevolent narrative, hero, is associated with an increase of 40%. The two narratives focusing on the malevolence of the enemy are instead less powerful (respectively, 36% for fool and 32% for villain), but still positive compared to the absence of a narrative.

On the negative end, users tend to retweet content from @uamemesforces and @DefenceU less than the content posted by the @Ukraine account (reference case). Further, the less popular content includes posters and illustrations, and those of scenes (as opposed to people or objects). However, one visual variable positively associated with retweets is whether it is a screenshot: these are often calls for donations and screenshots of news.

Overall, we find narratives to be a powerful predictor of the amount of attention a piece of content receives, while emotional and actor variables have a smaller impact.

4.3 Audience characterization

As summarized by the words of the creator of @Ukraine reported in Section 1, one of the main goals of Ukrainian memetic warfare is to reach “large and distant target audiences”, and thus influence global public opinion. In this section,

we study which countries are more receptive to Ukrainian messages, and how this receptiveness relates to narratives and to each country’s actions.

To examine the variability in responses in different countries, we look at retweets and geo-locate their authors to a country (see Section 3). We normalize the number of retweets in a given country by its population size, thus obtaining a measure of retweets per capita. This measure is a proxy for the amount of interest in Ukrainian tweets in a given country. Instead, to measure the support of a country to Ukraine in the conflict, we use economic and military assistance normalized by GDP, as explained in Section 3.

Figure 10 shows these two measures, retweet per capita and assistance, for each country present in our dataset. First, countries bordering Russia present a high retweet per capita (above 0.5‰): they pay increased attention to Ukrainian content. In particular, Baltic countries bordering Russia—Latvia, Lithuania, and Estonia—are outliers both in terms of assistance to Ukraine and of their engagement. Finland, also bordering Russia, presents a high retweet rate, albeit with a smaller aid statistic. Other European countries—especially northern (Norway, Denmark), eastern (Czechia, Poland), or English-speaking ones (Ireland, United Kingdom)—have high rates of retweets per capita (above 0.3‰).

The scatterplot shows a strong correlation between the two variables: with a Pearson coefficient of 𝜌 = 0.787, attention to Ukrainian content on Twitter is clearly related to the economic and military assistance that the country gives to Ukraine. While European countries bordering Russia are outliers, the correlation after excluding these countries is still a reasonably strong 𝜌 = 0.471 (𝑝 < 0.05, inset on the right). Thus, the relationship extends beyond countries bordering Russia. Conversely, the same analysis using the Eurobarometer survey shows no significant correlation between retweet per capita and popular support for financial aid to Ukrainian military efforts.

Let us now break down the geographic resonance of each narrative. To this end, we compute the odds of users in a country retweeting content with a particular narrative. An important working assumption in this analysis is that geolocation is not biased compared to the narrative that is shared, i.e., there is no reason to believe that a person who geolocates themself prefers one narrative over another. Figure 11 shows the retweet rate and the per-narrative log odds for each country in Europe. The focus on Europe derives from it being the geographic location of the conflict, and because it is the continent with the highest retweet rate per capita (0.3‰, compared to 0.09‰ in North America, which is second). The leftmost map shows the retweet rate normalized by population, which shows again the outsized attention by the Baltic countries and Finland. The retweets by narrative (rest of the plots) often present a gradient from West to East: countries closer to Russia are more likely to retweet posts with a discernible narrative. There are, however, some differences across narratives: the ‘victim’ narrative presents a more uniform geographical spread. In contrast, countries bordering Russia and Ukraine—such as Finland, Poland, and the Baltics—show a stronger resonance with the ‘villain’ narrative focusing on the Russian threat. In general, Baltic states are especially inclined to amplify

the archetypes of villain and fool, interestingly together with Greece (we may speculate that historical parallels with Cyprus’ experience of the contested occupation by Turkey may provoke attention from Greece). The ‘hero’ narrative appeals both to some Western countries—such as the United Kingdom, a strong supporter of Ukraine in the war—as well as those closer to the conflict, including Belarus. Note that Belarus has strong regulations on social media access, thus those who express themselves on social media are likely to be either those in opposition to the government and dedicated enough to use a VPN, or those who are allowed by the government on the site.[11]

Finally, we investigate how the receptiveness to different narratives is related to the actions of a country’s government in the conflict. Figure 12 shows the relationship between these odds and the total assistance to Ukraine. The total aid provided by each country to Ukraine, normalized by GDP, is positively associated with the spread of malevolent narratives (villain and fool) and negatively associated with benevolent ones (hero and victim). These relationships are not obvious, as tweets can be annotated with several narratives at once, for instance, the combination of hero and fool often appears in the data. We find that the above effect is mostly due to military aid, while humanitarian aid has no statistical relation to the resonance of narratives in our dataset (plots omitted for brevity). Again, a similar analysis shows no significant results for the Eurobarometer survey. These results echo the engagement model described earlier in that malevolent narratives are more positively associated with content virality. However, the engagement model still shows a small positive effect for the benevolent narratives, which instead show a negative relationship to aid. In other words, a benevolent narrative, such as ‘victim’, is likely to get more amplification. However, at the same time, malevolent narratives such as ‘villain’ are the ones positively associated with economic and military support.

[10] https://www.reuters.com/world/europe/russia-has-committed-wide-range-war-crimes-ukraine-un-inquiry-finds-2023-03-16

[11] https://www.cyberghostvpn.com/en_US/privacyhub/countries-ban-social-media