Changing Sources: Social Media Activity During Civil War

Anita Gohdes, Hertie School, Berlin and Zachary C. Steinert Threlkeld, University of California, Los Angeles[1]

Introduction

From isolated protests to country-wide uprisings or organized armed conflict, non-state and government actors have learned that their actions are likely to be caught on cellphone camera, and that controlling the digital narrative in the chaos of conflict can offer decisive advantages. Growing research has helped advance our understanding of how conflict shapes social media, and conversely, how social media influences conflict dynamics. Social media reduces the cost of communication,[2] increases the speed of its dissemination, and provides passive polling of conflict actors.[3] In addition, data gleaned from social media can be used to study network processes,[4] public opinion,[5] political representation,[6] protests,[7] and a wide range of other phenomena.[8]

In this essay, we investigate how social media usage may be influenced by local conflict dynamics. We study Twitter usage by individuals based in Syria during the conflict, including data from 2014 to 2017. Instead of studying the content posted by individuals using Twitter, we focus on account activity as an indicator of changing offline dynamics.  Narratives on social media may change not only because individuals change the type of content they post, but also because the composition of users posting from a certain location changes.

We compare the number of monthly accounts that were newly created as well as the number of newly inactive accounts in two areas with very different conflict dynamics: the rebel-dominated Jebel Saman district and the regime-controlled area of Latakia. Between 2014 and 2017, the Jebel Saman district, which includes the city of Aleppo, was subjected to heavy fighting, repeated changes in armed group presence, and was frequently the site of large government offenses against both civilians and opposition groups, all which resulted in high numbers of casualties.  Jebel Sama’s city Aleppo has been subject to some of the most intense fighting and changes in local control experienced throughout the period under investigation here.  The Latakia district in the northwest of the country includes the city of Latakia and is a pro-government stronghold, and is also home to bases for the Russian Navy and Air Force, the latter of which started engaging in the civil war in October 2015.  In both Latakia and the neighboring governorate of Tartus, the Alawi represent the majority of the local population. Since the beginning of the uprising, the Assad Regime has actively taken steps to promote sectarian divides, continuously emphasizing the regime’s close link to the Alawi community.[9]

We find that changes in conflict dynamics, such as the end of the regime-led siege in Aleppo in December 2016, coincide with substantial changes in local account composition. When scholars and practitioners rely on geo-located social media posts to make sense of a conflict, taking into account changes in the local composition of active accounts is crucial, but rarely done.

Despite significant insights into the use of social media by civilians proceeding political conflict, few studies examine civilians’ use of social media beyond initial mobilization. Research on wartime social media use by non-combatants has instead focused on issues such as  awareness raising campaigns, especially when they draw international attention to a subnational issue.[10] Analyzing the strategy of doctor-activists, for example, Alasaad (2013) shows how this specific group used Facebook and YouTube to spread international awareness about a leishmaniasis outbreak in Deir Ezzor province in 2013.[11]  In sum, the use of social media in the context of ongoing civil conflict, in particular when used by civilians, remains understudied and may be distorted by methodological issues, such as when neglecting changing patterns in users on the ground. In the following section we discuss social media usage during the Syria conflict and offer a descriptive analysis of activity patterns of Twitter users who identified their location in Syria between 2014 and 2017.

Social Media Activity During the Syrian Conflict

The Syrian conflict has been called the most socially mediated civil conflict in history,[12] with some commentators going so far as to claim that the internet itself has become a weapon of war.[13]  Armed actors on all sides of the conflict use social media to communicate their (change in) allegiances, spread propaganda, and interact with both their domestic and foreign audience.[14]  While major bans on social media platforms in Syria were lifted a few months prior to the outbreak of the conflict in 2011, Internet activity remains highly surveilled and controlled by the Syrian regime.[15] Countrywide Internet shutdowns have occurred numerous times, and the regime has strategically limited access to the Internet in certain governorates (administrative units equivalent to American states) as part of their broader repressive strategy.[16]

Because of the volume of social media content, and the relative inaccessibility of the country to researchers, Syria became a key testing ground for innovative methods for studying the conflict from afar.  Experts warned that the abundance of social media data provided observers only the illusion of complete information about events in Syria.[17]  But little research has been done on the effects of these often veiled limitations on or distortions of the data, or on the explicit ways in which dynamics of the conflict itself, such as changes in the composition of conflict parties and shifts in territorial control, directly impact the nature of social media discourse.

Measuring Account Activity

To understand how conflict affects social media behavior, we rely on geolocated Twitter posts. We collected the tweets in real-time, connecting to Twitter’s POST statuses/filter endpoint to collect those that include longitude and latitude coordinates. Globally, 2-3% of tweets contain location coordinates, though approximately 12.91% of Arabic tweets are geotagged.[18] Since Twitter  matches the parameters of a request up to a 1% ceiling, this process therefore provides between 7.7% ) and 50% ) of all tweets with coordinates.[19]  We query the stored tweets for those that were sent from Syria from April 2014 until October 2017. This filter provides 474,223 tweets from 20,926 unique accounts.[20] The following indicators are calculated at the monthly level:

  • Tweets: the number of tweets posted.
  • Active accounts: the number of unique accounts.
  • Account activity:
    • Accounts created: number of accounts created.
    • Accounts that go inactive: number of accounts that become inactive. An account is classified as inactive when it stops tweeting for at least three months. The start of inactivity is defined as the day on which they posted their last tweet.[21]

Figure 1: Activity by accounts geo-located in Syria, April 2014 – July 2017.

Note: The top panel shows the number of active accounts per month. The lower panel shows the number of tweets per month. After June 2015, they track each other much less closely than before.  Counts are at the country level.

Figure 1 shows the number the number of active accounts (top panel) and tweets (lower panel) by month from April 2014 until July 2017.[22]  The trends track each other before diverging.  From April 2014 to June 2015, the number of active accounts and tweets in Syria follow the same n-shape: increasing through October 2014, the number of active accounts and tweets steadily declines into June 2015. In October 2015, however, there is a surge of tweets without a corresponding increase in accounts. The sharp increase in posting activity coincides with the start of the military intervention of Russian forces in the conflict. In December 2016, there is a surge of account creation without a corresponding increase in tweets, and this increase corresponds with the end of the siege in Aleppo. Other than these two spikes, the trends match: from lows around June 2016, both slowly increase during the remainder of the sample.

To further investigate the trends in Figure 1, Figure 2 shows account activity in two districts of Syria, Jebel Saman and Latakia. In November 2016, regime forces had circled the remaining densely populated rebel-held areas in the Eastern part of Aleppo, and, in a coalition with the Russian Air Force, submitted the area to intense bombardment for twelve days, targeting core civilian infrastructure such as hospitals.[23] On December 13, a highly complex ceasefire was negotiated which involved the surrender of weapons and transfer of all remaining rebels to other territories, resulting in an estimated relocation of one hundred thousand individuals.[24] The relocation continued through December 15. The siege left the city destitute with tens of thousands dead, and many more close to starvation.

Figure 2 shows activity in Jebel Saman (left panels) and Latakia (right panels). The top panels show the monthly number of created accounts as well as the number of monthly accounts that become inactive. The lower panels show the monthly number tweets of during the same time period. These four panels show that district level trends sometimes match national ones and other times do not, and the times of divergence coincide with important offline events.

Before June 2015 the district-level trends are comparable to the national trends. In both districts, the n-shaped pattern reappears, including the steady drop through June. Jebel Saman’s tweets increase after June 15, but not consistently. In Jebel Saman, the descriptive graphs reveal intermittent spikes in tweet activity but not the steady increase seen at the country-level.  The graph shows that the end of the Aleppo siege was accompanied  by a significant change in the local composition of Twitter accounts. In the sample studied here, 369 accounts become inactive,  and  71 new ones are created. The large increase in inactive accounts in December 2016 starkly contrasts with country trends, which actually show an increase in the number of active accounts.

Figure 2: Changes in geo-located account activity, April 2014 – July 2017

Note: The top panel shows the number of accounts; the bottom shows the number of tweets. The left panel shows activity in Jebel Saman, an opposition stronghold, and Latakia, a pro-government district.  The spikes in Jebel Saman correspond to the end of the Aleppo siege, and the spike in Latakia occurs the same month the Russian Air Force, based in Latakia, intervened in the civil war.

An approximate estimate of the true number of accounts that go inactive is possible.  Using the percentage of Arabic users who geotag their tweet (12.91%) from Huang and Carley 2019 and assuming Twitter users from Syria geotag themselves at the same rate, then the true number of accounts that go active is approximately 2585 –  (.  The same calculation suggests that 549 () accounts activated at that time.

A closer look at the accounts that become inactive in Jebel Saman in December 2016 reveals similarities and differences to the overall sample of Syria-based accounts in the dataset. Qualitative studying of a random selection of account bios in both samples suggests similar types of accounts. In both the overall sample and the December 2016 Jebel Saman sample, only a small fraction of accounts tweet regularly (weekly). The accounts that go inactive in Jebel Saman have a higher average number of followers than the overall Syria sample, and post longer tweets (90 characters versus 68 characters). We also compare the sentiment of the tweets, counting the number of positive and negative words.[25]  Both samples use almost the same number of negative words, but the sample of accounts that go inactive in Jebel Saman use significantly fewer positive words than the Syrian sample does. Fewer positive words despite longer tweets suggests a less positive sentiment within the group of accounts that become inactive compared to those that remain active. Overall, accounts that go inactive are more popular and less positively valenced than those that stay active.

Looking at the new accounts created and active in Jebel Saman in December 2016, we observe that more of the Twitter bios are in English than in the overall sample. New accounts are roughly as active as the average account in the overall sample. Similar to the accounts that went inactive that month in Jebel Saman, the newly active accounts tweet longer messages (93 versus 68 characters). New accounts use comparatively fewer positive and negative words, despite posting longer tweets. Overall, newly active accounts seem to therefore convey less sentiment than the overall sample of tweets for Syria.

In Latakia, the pattern of activity in 2015 and 2016 looks very different than in Jebel Saman.  There is little to no variation in the number of accounts created or the number of accounts that cease activity. In contrast, posting activity rapidly decreased in the summer of 2015, only to then sharply increase in October of the same year. This rise in volume of tweets coincides with an overall spike in tweets geo-located in Syria in October 2015. While rises in Twitter activity in Jebel Saman – driven mainly by users based in Aleppo – are generally accompanied by a host of newly created accounts, changes in the number of tweets do not display this pattern in Latakia.

There are a number of possible explanations for these divergent patterns. Changes in posting frequency in Aleppo may be related to the death of individuals previously using Twitter, or to population movements within and outside of the area, either of which could contribute towards explaining the temporal variation in user composition. Alternatively, users in Aleppo may have chosen to abandon an old social media account and start a new one. Changes in account composition would then reflect changes in user behavior, rather than user composition. In light of the intense changes to the political and security context of the cases we studied, it is plausible to assume that the patterns of posting and account activity are a reflection of who was willing and able to be publicly active online. The indication of manipulated account activity at the outset of the Russian military involvement in the conflict further supports the argument that patterns of online activity are a function of changing conflict dynamics. We aim to further explore these dynamics in future work.

Finally, the rise in geo-located tweets in Latakia coincides with the overall activity visible in Figure 1 and the Russian Air Force’s involvement in the conflict in October 2015, and a handful of accounts almost exclusively drive this activity.  Moreover, these accounts appear to be manipulated. In our sample, 33 accounts tweeted from Latakia that month, but two authored 3,309 of all tweets; the third most active account only tweeted 156 times. These accounts present as a K-pop (Korean Pop Music) fan account, yet the second and third most active follow no accounts. In this paper’s sample, 89.67% of the three accounts’ tweets are from October 2015 and subsequent fan account activity is highly irregular – behavior consistent with manipulated accounts.  These accounts still exist, i.e. they have survived Twitter removal of coordinated account networks, though their (infrequent) tweets still focus on K-pop.

Following the taxonomy in Leber and Abrahams 2021, these accounts appear to be inauthentic and ambiguous.[26]  The irregular but heavy tweeting suggests a bot, but the coordinated nature of the activity and the apolitical activity are more suggestive of a coordinated support group. Official foreign influence operations these accounts are not.[27] We cannot identify the actor or actors managing the accounts, though other studies of digital influence operations in the region means Russia or a member of the KUBE block are the most likely culprits.[28],[29] Perhaps the most surprising feature of these accounts is that they do not appear to be part of a larger deception operation like often occurs in the region.  This apparent isolation could be because manipulated accounts rarely geotag tweets, so those identified here could be part of a larger campaign invisible to this paper’s sample.

Discussion

More than a decade after social media recorded and broadcast the first protests in the Middle East and North Africa, digital communication has become an everyday feature of modern conflicts. Much progress has been made in understanding the ways in which these tools are used for protest mobilization and coordination, yet beyond initial conflict onset, the everyday use of social media by civilians caught in the midst of war remains understudied.

While these results are descriptive, preliminary, and require further investigation, they emphasize that activity on social media reflects local changes in conflict dynamics. Shifts in territorial control at the local level may significantly change the composition of local users who are active on social media, resulting in shifting patterns of content that are attributable to shifting activity patterns, and not necessarily shifting sentiment of emotion of the same individuals. In future iterations, we will continue to build knowledge about how conflict dynamics affect social media usage.

 

 

[1] We thank Lucien Baumgartner for excellent research assistance.

[2] Andrew T Little, “Communication Technology and Protest,” Journal of Politics 78, no. 1 (2015): 152–166.

[3] Thomas Zeitzoff, “How Social Media Is Changing Conflict,” Journal of Conflict Resolution 61, no. 9 (2017): 1970–1991, doi:10.1177/0022002717721392.

[4] Daniel M Romero, Brendan Meeder, and Jon Kleinberg, “Differences in the Mechanics of Information Diffusion Across Topics: Idioms, Political Hashtags,  and Complex Contagion on Twitter,” in  Proceedings  of  the  20th  International  Conference  on  World  Wide  Web  (ACM, 2011), 695–704, isbn: 9781450306324;  Pablo Barberá et al., “The Critical Periphery in the Growth  of Social Protests,” PloS ONE 10, no. 11 (2015): 1–15, doi:10.7910/DVN/WCXK3Z.Funding, https://doi.org/10.1371/journal. pone.0143611.

[5] Thomas Zeitzoff, “Does Social Media Influence Conflict? Evidence from the 2012 Gaza Conflict,” Journal of Conflict Resolution 62, no. 1 (2018): 29–63, doi:10.1177/0022002716650925; Nicholas Beauchamp, “Predicting and Interpolating State- Level Polls Using Twitter Textual Data 00 State-level public,” American Journal of Political Science 61, no. 2 (2019): 490–503, doi:10.1111/ajps.

[6] Stan Oklobdzija, “Dark Parties: Citizens United, Independent-Expenditure Networks and the Evolution of Political Parties,” in Political Networks Workshops & Conference (2018); Barberá, Pablo, Casas, Andreu, Nagler, Jonathan, Egan, Patrick J., Bonneau, Richard, Jost, John T., & Tucker, Joshua. (2019). Who Leads? Who Follows? Measuring Issue Attention and Agenda Setting by Legislators and the Mass Public Using Social Media Data. American Political Science Review, 113 no. 4 (2019), 883-901. doi:10.1017/S0003055419000352; Pablo Barberá and Thomas Zeitzoff, “The New Public Address System: Why Do World Leaders Adopt Social Media?,” International Studies Quarterly 62, no. 1 (March 2018): 121–130, doi:10.1093/isq/sqx047.

[7] Marlon Mooijman et al., “Moralization in social networks and the emergence of violence during protests,” Nature Human Behaviour  2,  no.  June  (2018):  389–396, doi:10.1038/s41562- 018- 0353- 0,  http://dx.doi.org/10.1038/ s41562-018-0353-0; Jennifer M Larson et al., “Social Networks and Protest Participation: Evidence from 130 Million Twitter Users,” American Journal of Political Science 63, no. 3 (2019): 690–705, doi:10.1111/ajps.12436; Anton Sobolev et al., “News and Geolocated Social Media Accurately Measure Protest Size” (2020).

[8] Pablo Barbera and Zachary C Steinert-Threlkeld, “How to Use Social Media Data for Political Science Research,” chap. 23 in  The  SAGE  Handbook  of  Research  Methods  in  Political  Science  and  International  Relations  (London:  SAGE  Publications Ltd, 2020), 404–424.

[9] See e.g. Michael Kerr and Craig Larkin, The Alawis of Syria: War, Faith, and Politics in the Levant (New York: Oxford University Press, 2015).

[10] Abeer Najjar, “Othering the Self: Palestinians Narrating the War on Gaza in the Social Media,” Journal of Middle East Media 6, no. 1 (2010): 1–30.

[11] Samer Alasaad, “War diseases revealed by the social media: massive leishmaniasis outbreak in the Syrian Spring,” Parasites & vectors 6, no. 1 (2013): 1–3.

[12] Marc Lynch, Deen Freelon, and Sean Aday, “Blogs and Bullets III: Syria’s Social Mediated War,” United States Institute of Peace, Peaceworks 91 (2014): 5.

[13] Mohamed Hashem, “Q&A: In Syria the ’internet has become a weapon’ of war,” Al-Jazeera, 2015, http://www.aljazeera. com/indepth/features/2015/06/qa-syria-internet-weapon-war-150619215453906.html.

[14] Dana M Moss, “The ties that bind: Internet communication technologies, networked authoritarianism, and ‘voice’ in the Syrian diaspora,” Globalizations  15, no. 2 (2018): 265–282, doi:10.1080/14747731.2016.1263079, https://doi.org/10.1080/14747731.2016.1263079.

[15] Freedom  House,  “Syria,”  Freedom  on  the  Net  2015,  2015,  https://freedomhouse.org/sites/default/files/resources/ FOTN%202015_Syria.pdf.

[16] Anita R Gohdes. “Repression Technology: Internet Accessibility and State Violence.” American Journal of Political Science, 64:3 (2020):s 488-503. https://doi.org/10.1111/ajps.12509

[17] Lynch, Freelon, and Aday, “Blogs and Bullets III: Syria’s Social Mediated War”; Megan Price, Anita Gohdes, and Patrick Ball, “Documents of war: Understanding the Syrian conflict,” Significance 12, no. 2 (2015): 14–19, doi:10.1111/j.1740-9713.2015.00811.x.

[18] Binxuan Huang and Kathleen M. Carley.  “A Large-Scale Empirical Study of Geotagging Behavior on Twitter”.  Proceedings of the 2019 IEEE/ACM International Conference on Advances in Social Network Analysis and Mining.  2019: 365-373.

[19] For  more information on working with Twitter,  see Zachary C. Steinert-Threlkeld (2018). Twitter as Data (Elements in Quantitative and Computational Methods for the Social Sciences). Cambridge: Cambridge University Press.

[20] The data collection process failed to collect tweets for December 2015.

[21] We choose to measure account inactivity retroactively through the date of the last tweet  because  actually  deleting  an account can pose difficulty, in particular when accessing Twitter from a mobile device or through an app. We assume that users who either switch accounts or stop using their account will more likely stop actively using it long before they eventually close it (if at all). Users may sign up for a new account, and access this through the Twitter app without actually deleting their old account simultaneously. In addition, Twitter’s API only confirms that an account no longer exists but not its date of cessation.

[22] We restrict the figures to three months before the end of our observation period to allow for the classification of inactive accounts.

[23] Annabelle Böttcher, “News Analysis Humanitarian Aid and the Battle of Aleppo,” 105, no. 1 (2017): 2-3.

[24] Laila Bassam, Angus McDowall, and Stephanie Nebehay, Battle of Aleppo ends after years of bloodshed with rebel with- drawal,  2016,  https://www.reuters.com/article/us- mideast- crisis- syria/battle- of- aleppo- ends- after- years- of- bloodshed-with-rebel-withdrawal-idUSKBN1420H5; Atlantic Council, “Breaking Aleppo,” The Atlantic Council of the United States, 2017, 1–70.

[25] We split the corpus by language and apply separate, translated lexicons to each using the R package syuzhet. Saif M Mohammad and Peter D Turney, “Emotions Evoked by  Common Words  and Phrases: Using Mechanical Turk  to Create  an Emotion Lexicon,” in Proceedings of the  NAACL  HLT  2010  Workshop  on  Computational  Approaches  to  Analysis  and Generation of Emotion in Text, CAAGET ’10 (USA: Association for Computational Linguistics, 2010), 26–34. We use the get_nrc_sentiment function for English and Turkish Tweets. For Arabic tweets we replicate the procedure using the Arabic translation of the NRC Emotion Lexicon, created by Mohammad Salameh, Saif M. Mohammad, and Svetlana Kiritchenko http://saifmohammad.com/WebDocs/Arabic%20Lexicons/nrc_emotion_ar.txt

[26] Andrew Leber and Alexei Abrahams.  Social media manipulation in the MENA: Inauthenticity, Inequality, and Insecurity.  POMEPS Special Series 2021.

[27] Alexandra A Siegel.  Official Foreign Influence Operations: International Broadcasters in the Arab Online Sphere.  POMEPS Studies 43. (2021).

[28] Akin Unver and Ahmet Kurnaz.  Russian Digital Influence Operations in Turkey 2015-2020.  POMEPS Studies 43. (2021).

[29] Marc Owen Jones.  Tracking Adversaries: The Evolution of Deception and Manipulation Tactics on Gulf Twitter.  POMEPS Studies 43. (2021).