Hanna Hosman, Joost Ingen-Housz, Henna Paakki, Carlos Rosas
Our analysis of a mixed group of far-left Latin American actors linked with a far-right Trump supporting US based cluster of Facebook pages and groups shows that these ideologically different clusters were united in their mobilization of anti-NATO narratives. They are linked through link sharing behavior via a Facebook page called Bible Prophecy in the news, a religious cult sharing attention-grabbing content and disinformation about the Ukraine war. We found that the main narratives pushed by this network of actors tended to concentrate on societal topics that are prevalent and news-worthy globally as well as socially divisive – meanwhile having potential for anti-NATO sentiment. Before the Ukraine invasion, these included critical discussions around the Cuban embargo, anti-government or corruption narratives, and vaccines. However, starting from the Ukraine invasion, the focus of the anti-NATO narratives shifted their focus on topic related to the Ukraine war, anti-Ukraine and pro-Putin stories, digressive topics and disinformation content.
The Russian-Ukrainian conflict has been developing not only on the military front, but also on other fronts such as international relations, economy and information sharing. Public opinion regarding the actions of the actors involved can play a crucial role in the conflict, since it will determine the approval or disapproval of the political leaders and as a consequence the role they choose to play in the war.
In this sense, it is no secret that for years private and public actors have been developing large campaigns in the field of social networks, with the aim of influencing public opinion, with resources that can range from massive advertising to what will be our object of study: coordinated link sharing behavior (CLSB) of problematic information about the war.
Using the dataset built by Fabio Giglietto and his team, we will focus on the analysis of clusters of Latin American Facebook groups, where CLSB has been detected. This choice seems to us really interesting, since one could think that Latin America is an actor with little or no relation to a conflict between European countries, but the analysis we will perform will show us that this is not the case and that these CLSB campaigns directed towards Latin America have a narrative-building motive.
We used a mixed approach to study this linked component more closely. We analyzed all coordinated links shared by groups and pages included in this component, one year before the Ukraine invasion and one year after. Utilizing both computational Natural Language Processing (NLP) and topic modeling methods, as well as network analysis, guided by and enhanced with qualitative close-readings of most influential accounts and actors, and content shared, we strived to form an understanding of the main narratives pushed by this network of actors.
Our research questions were:
Did the narratives that the groups pushed change when the attack against Ukraine started?
Does the change reflect some specific goal?
Why is a far-left group of clusters connected with a far-right USA cluster?
Using the Crowdtangle-extracted entire dataset on coordinated link sharing behavior by Giglietto et al., we examined the network of pages that shared links in a coordinated manner by using Gephi. In the Gephi network graph, we identified a component of linked Facebook pages which was formed of four linked clusters of facebook pages and groups: a Mexican cluster, Argentinian cluster, Peruvian cluster, and a US-based cluster. Filtering the data further, we examined this particular component in more detail, identifying that the US cluster was linked to the South American clusters through a group called “Bible prophecy in the news”. To further study this component, we scraped all coordinated link sharing of the pages and groups that belonged in this component from a one year period before the Ukraine invasion, and one year after, using Coornet and Crowdtangle. We examined the content shared using mixed methods: computational analyses including network analysis, BERT topic modeling and Natural Language Processing (NLP), and qualitative close-reading and analysis of most important contents.
Firstly, we examined the component more closely by creating a network of links that were most connected to the clusters within our component, with frequency of link shared as edge weights. We also examined the domains the clusters most frequently interacted with using the same network analysis approach. For both component-centric networks we used Gephi for visualization and analysis of the networks.
Secondly, we used BERT topic modeling (Grootendorst, 2022), the multilingual version, to identify the most frequent topics discussed during both periods that spanned our data: period 1) representing the 10 months before the Ukraine invasion, and period 2) representing the 10 months starting from and after the invasion. We extracted the most central topic terms that differed among periods 1 and 2 (Cuba, bloqueo, corrupcion, vacuna, guerra, Malvinas, nazi, Putin, oro, Rusia, Ucrania). We then compared the usage frequency of these topic terms during these two periods using Iramuteq’s (http://www.iramuteq.org/) term frequency analysis function.
We further analyzed the second period discourses on the Ukraine invasion by using NLP methods. We used python scripts to extract all Facebook post titles, message texts, image text and post descriptions from the Crowdtangle data from all clusters in our component, from all shared posts that were coordinatedly shared during period 2. We combined these into one text for each post, and with NLTK libraries (Bird et al., 2009) preprocessed the texts to remove punctuation and to tokenize the texts, NLTK stopword corpora for English and Spanish to remove stopwords, and simplemma (Pypi library) to lemmatize the Spanish language texts. Then we used sklearn (Pedregosa et al., 2011) feature extraction library’s CountVectorizer for calculating a word co-occurrence matrix for the texts. After extracting a highly sparse matrix of co-occurrences, we created an edge list dataframe with “Source”, “Target” and “Weight” columns, required to build a network of word co-occurrences, based on the occurrence matrix. We then further filtered and reduced the size of the edge list by calculating the the mean of all weights and their standard deviation, and then including in the final edge list only word combinations with weights that exceeded a limit of 2*SD over mean weight. This resulted in a final edge list of 57,782 rows. Finally, we imported the edge list to Gephi for further analysis, and filtered the network to show only words that were in the network connected to the term Ucrania (Ukraine).
Finally, we conducted qualitative analysis of most central groups and pages involved in coordinated link sharing, the most shared posts, and the accounts and groups which had gotten most interactions on social media. This involved investigating the owners of the most central groups, the owners’ and the groups’ political and ideological positions and views and narratives on Ukraine invasion, the messages and ideological positions conveyed in the most shared or most interacted with posts, and what was shared between clusters through the “Bible prophecy in the news” group that linked South American clusters with the USA cluster.
Figure 1. Component 10 Network (4 clusters)
Each of the group members focused on the qualitative exploration of one of the clusters: reading and interpreting the group's names, the publications with the most interactions and analyzing what were the main topics of discussion.
What we found was that the three South American clusters were heavily left to extreme left oriented. These clusters are mainly formed of groups that propose as main topic the support or defence of some representative left-wing political figure or the direct opposition to political parties considered as neoliberal or right-wing (in the case of Mexico we observed mainly support to the current president Lopez Obrador, in the Peruvian case to the recently dismissed Pedro Castillo and in the Argentinean case we observed several groups linked to the Kirchnerist movement).
However, the most interesting feature that we found was that these three Latin American clusters linked to a US cluster – moreover, because the analysis revealed that the themes of the groups that make up the North American cluster is totally opposite to that of the other three: the vast majority of groups are pro-Trump or support a public figure from the US right wing political front. The Facebook group that serves as a link between the Latin American clusters and the US cluster is a group called "Bible Prophecy In the News". At first glance, this group appears to be a group dedicated exclusively to religious content, but a review of the videos and images posted allows us to observe that the speakers who appear in these publications also comment on global political news, mentioning the Russian-Ukrainian conflict as a supposed example of "the second coming of Christ and the end times".
Discussion with our team of project tutors about this group revealed to us that this group and its publications showed very similar aesthetic and discourse characteristics to those of the Chinese religious sect "Eastern Lightning". This religious sect, which has been classified by the Chinese government as a terrorist group, usually uses as a recruitment tool the viral sharing in various social networking groups of shocking videos or images about current popular events, such as alleged deadly effects of the vaccine against Covid-19 or in our particular case, news about the Russia-Ukraine conflict. Once a user becomes interested in those live seminars or YouTube channels and contacts the creators, Eastern Lightning uses this opportunity to begin to gradually introduce him or her to the cult. These statements are reinforced after verifying the history of name changes of this group (more than 15 in 3 years, including "COVID 19 PANDEMIC").
On the other hand, knowing that the general themes of the three Latin American clusters are linked to the left and those of the US group are related to political right, the analysis of the particular topics of discussion reveals an interesting fact: the narratives they tell frequently seek to build anti-US or more specifically anti-NATO sentiment.
For example, in the case of the three Spanish-speaking clusters, we can easily observe that the anti-right publications present almost always end up blaming the United States for having manipulated or influenced the situation, qualifying the local politicians as mere lackeys of these interests. In the case of the US right-wing cluster, the attacks are not positioned against the country itself, but rather against NATO and the public figure of Joe Biden, who according to them is nothing more than a puppet who allows himself to be led by the pressures of other countries that make up this organization.
The same logic is applied to the narratives about the conflict: according to these publications, the conflict is a great conspiracy that would allow the United States and NATO to have an excuse to invade Russia and overthrow Putin, who is the only one who currently maintains a balance of power against these actors. According to these narratives, it would be most convenient for Latin America to align itself with Russia in order to fight against the United States and its interference.
Figure 2. Term co-occurrence in the Latin American clusters' Ukraine war related publications.
Figure 3. Term usage variation 10 months before and 10 months after war start (Latinamerican clusters).
On the other hand, what we found by conducting the word co-occurrence analysis regarding texts related to Ukraine narratives is the presence of many terms that are part of popular disinformation narratives about Ukraine and Zelensky. The most notable cases are the comparison of the Russian invasion to the United Kingdom intervention on the Falkland Islands (Malvinas), the narrative about Zelensky and Ukraine actually using the ressources other countries lent them in order to buy and hoard gold (oro) or the one about the necessity of Russian intervention due to neonazi presence in Ukraine.
However, our most interesting finding was a narrative shift between the two periods of analysis (10 months before and 10 months after the invasion). The main topics of discussion present in Latin American clusters before the invasion (government corruption, vaccination, Cuban embargo, etc) seem to have almost vanished in the second period, replaced by the war and disinformation related narratives and topics. Of course, an attention shift is expected when a global scale event occurs, but what we appreciate here is not a diminution of the discussion about the pre-invasion topics, but almost a complete disparition. Furthermore, what makes the shift interesting is the form that the post-invasion narratives take, as discussed thus far, pushing disinformation narratives and an anti-NATO ideology to affect public opinion on social media.
Based on our findings, we can affirm that the narratives shared by the clusters analyzed changed after the beginning of the war. As we commented above, this change occurred in a rapid and extensive manner, resulting in the disappearance of the old narratives. Likewise, this rapid change suggests the former presence of networks of coordinated news and links sharing on specific topics, which have the capacity to change the main narrative shared according to global events to suit their own agenda.
So, what could be the main objective of this coordinated sharing? If we evaluate the pre-invasion and post-invasion narratives, though they seem to have little in common (narratives on Covid-19 vaccines vs. narratives on neo-Nazis in Ukraine), they all lead to the same shared agenda: the construction of anti-US or anti-NATO sentiment. The amplification of such sentiment clearly benefits Russia, especially if it occurs in geographic zones that could be considered as "neutral" in this global power struggle, considering the important role that Latin America can play in the next 20 years due to its large amount of natural resources such as minerals and water. Likewise, anti-US sentiments seem to be on the rise in Latin America during the last 5 years, a reality that is reflected in the election of leftist presidents in several Latin American countries (Boric in Chile, Castillo in Peru, Petro in Colombia, Lula in Brazil, AMLO in Mexico, etc).
On the other hand, the fact that the rapid variation of topics has been accepted by the members of these groups, who continue to receive comments and interactions, suggests that a good part of these members do not join these groups in support of a specific cause or to protest against it, but rather an anti-system, anti-establishment attitude, which seeks to profit from the current events to express their disagreement and rejection of the current order of things. In this sense, variations of the group named "Bible Prophecy in the News" that we mentioned earlier are not uncommon in the make-up of the social media clusters we have analyzed, especially if they are groups that were formed in support of a particular public figure and that figure is no longer popular or has fallen out of favour.
We examined a mixed network of far-left Latin American and far-right USA groups linked through coordinated link sharing behaviour via a Facebook page called Bible Prophecy in the news. Our analysis shows that although these groups are political opposites, they are united in their anti-establishment and anti-NATO ideology, sharing content relevant to these ideological interests. The popular feelings of dissatisfaction with the system that are visible in this network can be exploited by public opinion manipulation networks such as the CLSB networks we have observed here. These networks find fertile ground in the members of these clusters, who do not have much difficulty in following issues as long as they allow them to express their anti-establishment ideology. This ideological leaning is also present in the connection between the US cluster and the Latin American clusters, regardless of the fact that the political positions of these clusters are totally opposite, as both can find a common ground in the expression of their support for anti-establishment sentiment. This involves utilising globally relevant events like the Ukraine war as vehicles for backing their agenda.
Although the existence of such disinformation networks has been uncovered by earlier research on disinformation (Badawy et al., 2018; Bennett and Livingston, 2018), future research should investigate in more detail the spread and influence of such campaigns, and variety and scope in how they might affect public opinion, e.g. in populations that are undecided or in between i.e. not in the extreme in their political ideology.
Another interesting approach that we did not have time to explore would be to do a qualitative check for the possible presence of pro-Ukrainian disinformation link sharing. During the conflict, Ukraine has been backed by the United States, a country also known to have applied massive propaganda and disinformation dissemination tactics in the past (e.g. Humprecht, 2018). Therefore, a scan of Facebook with CrowdTangle using another set of keywords could reveal the presence of other types of coordinated link-sharing networks.
Finally, we believe that interesting findings could emerge if we apply the same methods we have used to all the other clusters present in the network, especially regarding the groups serving as links between two opposing political clusters, which could give us a clue as to whether the anti-establishment sentiment found in our component could be found in other parts of the entire network as well.
We used the Crowdtangle-extracted dataset on coordinated link sharing behavior by Giglietto et al., which was originally found cased on a key word search related to the Ukraine war.
Second, after identifying the interesting groups in the whole network of actors, component 10, we scraped all coordinated link sharing of the pages and groups that belonged in this component from a one-year-period before the Ukraine invasion, and one year after, using Coornet and Crowdtangle.
Badawy, A. and Ferrara, E. and Lerman, K.. 2018. Analyzing the Digital Traces of Political Manipulation: The 2016 Russian Interference Twitter Campaign. In 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM). IEEE, Barcelona, 258–265. https://doi.org/10.1109/ASONAM.2018.8508646
Bennett L. and Livingston, S. 2018. The disinformation order: Disruptive communication and the decline of democratic institutions. European Journal of Communication 33, 2 (April 2018), 122–139. https://doi.org/10.1177/0267323118760317
Bird, S. and Klein, E. and Loper, E. (2009). Natural Language Processing with Python. O’Reilly Media Inc. https://www.nltk.org/book/
Grootendorst, M. (2022). BERTopic: Neural topic modeling with a class-based TF-IDF procedure, arXiv preprint arXiv:2203.05794. https://github.com/MaartenGr/BERTopic
Humprecht, E. (2019). Where ‘fake news’ flourishes: a comparison across four Western democracies. Information, Communication & Society, 22(13), 1973-1988. DOI:10.1080/1369118X.2018.1474241
Pedregosa, F. and Varoquaux, G. and Gramfort, A. and Michel, V. and Thirion, B. and Grisel, O. and Blondel, M. and Prettenhofer, P. and Weiss, R. and Dubourg, V. and Vanderplas, J. and Passos, A. and Cournapeau, D. and Brucher, M. and Perrot, M. and Duchesnay, E. (2011). Scikit-learn: Machine Learning in Python, Journal of Machine Learning Research (JMLR), 12, pp. 2825-2830.