The affiliative use of emoji and hashtags in the Black Lives Matter movement: A Twitter case study

Protests and counter-protests seek to draw and direct attention and concern with confronting images and slogans (1–3). In recent years, as protests and counter-protests have partially migrated to the digital space, such images and slogans have also gone online (4–6). Two main ways in which these images and slogans are translated to the online space is through the use of emoji and hashtags. Despite sustained academic interest in online protests (7–9), hashtag activism (10–12) and the use of emoji across social media platforms (13–15), little is known about the specific functional role that emoji and hashtags play in online social movements. In an effort to fill this gap, the current paper studies both hashtags and emoji in the context of the Twitter discourse around the Black Lives Matter movement. 1


Introduction
P rotests and counter-protests have long made effective use of images and slogans (1)(2)(3).As protests and counterprotests have partially migrated to the digital space, such images and slogans have also gone online (4)(5)(6).Two important ways in which these images and slogans appear is through the use of emoji and hashtags.Some emoji have readily identifiable offline counterpartssuch as the raised fist, which was first deployed as a standalone image in protests in the San Francisco Bay Area and Harvard University in 1968-9 (1), and which now has its own emoji.
Similarly, some hashtags (like #BlackLivesMatter) reflect wellknown offline slogans.Indeed, on Twitter since 2016, this hashtag is automatically enhanced with a small emoji-like sticker featuring a trio of raised black, brown, and white fists.
The use of other emoji and hashtags, however, can be more obscure.To shed some light on their functions, we here study emoji and hashtags embedded in tweets associated with the Black Lives Matter protests in 2020, including the right-wing backlash to those protests.We analyze a dataset covering both the lead-up to and the aftermath of the 25 May 2020 murder of George Floyd by Officer Derek Chauvin in Minneapolis.
The nine-minute video of Floyd's murder set off a firestorm of activity both in the streets and online (see also SI Appendix).
While the use of hashtags as an organizing mechanism in online activism has been studied (10), the role of emoji in social movements has, to the best of our knowledge, received no academic attention.At the same time, as emoji have become an increasingly popular form of communication, a growing body of work that tracks the various types and uses of emoji has emerged (13)(14)(15).Extrapolating from this literature, we present and test four hypotheses regarding the use of emoji in online activism.First, emoji might be used for their straightforwardly semantic content, functioning as 34 compact logograms that efficiently convey meaning within the 35 tight character constraints of Twitter (H1).Second, emoji 36 and hashtags might be employed to disambiguate tone in the 37 context of highly-charged discursive exchanges (H2).This 38 follows from the observation that emoji and hashtags enable 39 us to track important linguistic subtleties-such as sarcasm 40 and humour-that are otherwise hard to detect in computer-41 mediated communication (16)(17)(18).Third, emoji might operate 42 on par with ostensive interlocutory gestures that aim at draw-43 ing and directing attention to the content of a given message

Significance Statement
Protests and counter-protests have partially migrated to the digital space.In so doing, images and slogans that would otherwise have appeared on protest signs, posters, and t-shirts have been replaced by emoji and hashtags.Despite sustained academic interest in hashtag activism and a growing body of work on the use of emoji across social media platforms, little is known about the functional role that emoji and hashtags play in online social movements.The current paper bridges this gap by identifying and interpreting the ways in which communities involved in the Twitter discourse around the Black Lives Matter movement employed emoji and hashtags in 2020.Utilizing mixed methods, we reach two important conclusions.First, distinct online groups employ distinctive patterns of emoji and hashtags.The observed patterns in emoji and hashtag use are both informative and interpretable, which offers an effective and computationally efficient way to make sense of online social movements.Second, emoji and hashtag use appears to play a complex role in the attention economy.Emoji and hashtag use tends to decrease engagement with individual tweets, but increase the attention paid to the same author on other tweets.This suggests a primarily affiliative use of hashtags and emoji: they are used to signal ingroup loyalty, at the expense of immediate attention.

D R A F T
their group (H4).This fourth function is especially relevant the 100 most-influential nodes within each group.Based on this, we characterize the four communities as follows.
Activists: this cluster represents the core of the movement and reflects the grass-roots nature of Back Lives Matter.It features a heterogeneous collection of individual activists, many of whom explicitly endorse the movement by placing #BlackLivesMatter in their profile bio.
Progressives: this cluster contains a range of high-profile individuals and organizations that are generally supportive of the Black Lives Matter movement.In addition to prominent Democratic politicians (e.g., Kamala Harris, Bernie Sanders) and liberal media outlets (e.g., ABC, NBC, CBS), there are various non-profit organizations and legal aid foundations (e.g., ACLU).
Reactionaries: this cluster features conservative politicians and public figures (e.g., Donald Trump, James Woods), as well as right-leaning media outlets (e.g., Breitbart News, The Gateway Pundit), anti-BLM activists, supporters of the police, and a large number of conspiratorially-minded and openly racist individuals.Additionally, this community features an exceptionally large number of suspended accounts.Though we cannot interpret these accounts, it stands to reason that they violated Twitter's community guidelines regarding false (conspiratorial) and offensive (xenophobic) content.
Boosters: this cluster comprises a diverse collection of individuals whose primary involvement with the Black Lives Matter movement seems to consist of link-sharing and fundraising.With respect to fundraising, we identify a large contingent of fans of the Korean pop phenomenon, commonly called K-pop.While the link between K-pop and Black Lives Matter might seem tenuous at first, the fact that one such band-Bangtan Sonyeondan (BTS)-donated a million dollars to the Black Lives Matter foundation, and encouraged its followers to match that donation, explains this group's engagement (28).
Based on these initial impressions, and in effort to test whether these different communities use emoji and hashtags in differential ways, we next constructed a classifier.
Classifier Construction.The prospect of using emoji to train classifiers has received considerable attention and produced impressive results (13,29).Classifiers are able to disambiguate the tone and intention of a given statement by differentiating between positive and negative valences of the same emoji (30)(31)(32).Extending this line of research, we investigated RQ1 by constructing a classifier of our own, with the goal of using emoji and hashtags for community detection.The goal was to provide a measurement of the linguistic cogency of the communities we identified through modularity analysis, and to provide a measure of entropy that shows how much additional information about community membership is contained within each type of communication.
We constructed a standard data-science workflow in Python for automated text classification, using Tensorflow (33) for neural network approaches and Scikit-Learn (34) for classical classification techniques.
Excluding members of the Boosters cluster because there were too few of them for meaningful analysis, we generated a document containing the plain text of all tweets by each user in the network, together with the emoji and hashtags they used.Tweet text was preprocessed using a typical text processing workflow (removing nonalphanumeric, non-hash identifier (#), and non-emoji characters, standardizing case, etc.), and emoji were encoded using Python's emoji package (35).Further details on the data sampling, train/test splits, as well as evaluation, are available in the SI Appendix.

Distinctiveness.
To examine the distinctiveness of each of the communities, we first determined the 15 most popular emoji for each community c and then for each emoji e.From that set we calculated P r(c|e).This gives a rough measure of how much information emoji and hashtag use carries about community membership, as well as which specific emoji and hashtags are most distinctive within each community.

Emoji and Hashtag Co-occurrence Network Construction. Though
informative in its own right, we noted that a single tweet may well contain multiple emoji, multiple hashtags, and various combinations thereof.On this score, previous research has found that multiple hashtags are often combined to draw attention to interrelated issues D R A F T (23,36).Likewise, various emoji are frequently used together to refine a user's stance, attitude, or sentiment (13-15)(see also SI Appendix).Accordingly, it is informative to consider whether and in what ways the various communities combine emoji and hashtags.
To this end, we conducted a co-occurrence analysis.
For each of the four communities identified by the social network analysis, we identified the fifty most-commonly-used emoji and the fifty most-commonly-used hashtags.We then constructed a co-occurrence network of these emoji and hashtags for each community, where the nodes are either emoji or hashtags and the edges between them represent co-occurrence in the same tweet.These networks enable us to answer the question, "What do people (in this community) talk about, and how do they talk about it, when talking about X?" We then visualized these networks using Gephi (37) and the Image Preview plugin (38) to provide a snapshot of the imagery and slogans most distinctive of each community.

RESULTS
Community Structure.As Figure 1 shows, the retweet network is bipolar, with Activists, Progressives, and Boosters on one side and Reactionaries on the other.This outcome is consistent with numerous findings from political science suggesting substantial polarisation in the political landscape (39,40).
As Figure 2 shows, the murder of George Floyd triggered an outpouring of tweets -first among Activists, then among Progressives and Boosters, and finally among Reactionaries.
The decay in the volume of tweets among these groups is also worth considering, as Reactionaries decay much less quickly than the three other communities, suggesting a self-sustaining dynamic within that community.In contrast to both Activists and Progressives, Reactionaries (Figure 4c) did not appear to center their attention on any single topic.Instead, different elements of this group pushed back against the Black Lives Matter movement in different ways.For instance, we see a large contingent drawing attention to the police via both emoji (blue heart �, police officer � � � � � � � � � � � � � �, police cruiser � � � � � � � � �) and hashtags (e.g., #backtheblue, #bluelivesmatter), while other tweets seem to focus more on electoral politics, either by identifying with the Trump reelection campaign (e.g., #kag2020, #trump2020) or by derogating enemies (e.g., #democratsaredestroyingamerica, #liberalismisamentaldisorder).In addition, we see the centrality of the QAnon conspiracy theory to this community in its use of hashtags such as #qanon, #wwg1wga ("where we go one, we go all," a popular slogan in the QAnon movement), and #thegreatawakening.The Reactionary community does not seem to unite around a single cause or message; instead, they are primarily defined in terms of what they oppose.It appears that this reactionary movement in part reflects an attempt to hijack the Black Lives Matter discourse in order to bootstrap its own political agenda (42).
It is also worth remarking that Progressives and Reactionaries used distinctive hashtags to collate tweets about the protests in Seattle, Washington: Progressives employed #seattleprotest and #seattleprotests, whereas Reactionaries used #chop and #chaz (referring to the anarchist zone that protesters set up on 8 June 2020, and which the Seattle Police Department cleared on 1 July 2021 after an unlawful assembly was declared).This suggests a potential filter bubble effect, in which users who followed one set of hashtags about the Seattle protests would encounter the corresponding set of information and sentiment about it, while users who followed the other hashtags would encounter a radically different set of information and sentiment about the same topic.
Finally, the Booster community (Figure 4d) exhibits many D R A F T  use are given in Table 2. Across communities, about 25% of users have at least one emoji in one of their tweets.Hashtag usage is much higher, averaging about 57% but ranging as high as 90% in the Booster group.For all communities, both emoji-and hashtag-users have a higher PageRank, suggesting that they are more embedded in the network as a whole.
Retweet statistics show several curious patterns for both emoji (Table 3) and hashtag (Table 4) usage.For ease of interpretation, we divide these into three patterns: Type I patterns: when both the following conditions are met: Type III patterns: anything not in Types I and II.
Note that, with two exceptions, both the overall and community-level patterns fall into either Type I or Type II.
That is, in general, EH-usage does not provide a substantial  advantage in mean retweets, and often gives a substantial dis-360 advantage, compared to tweets by people who never use them.361 However, EH-users' other tweets are retweeted on average 362 much more than non EH-users.

363
The two Type-III exceptions are found in Reactionaries' 364 use of emoji (which appears to make no particular difference 365 to retweets) and Boosters' use of emoji (which overall confers 366 substantial advantages and may reflect the multilingual nature 367 of this community and the fact that emoji can serve as a sort 368 of lingua franca).With respect to RQ1, Table 1, Figure 3, and Figure 4 provide 371 compelling evidence that there are informative and meaningful 372 differences in the way that the various communities involved in 373 the Black Lives Matter discourse employ emoji and hashtags.374 The results of our classification task confirm that there are 375 distinct patterns in emoji and hashtag usage across the various 376 communities and that these differences are informative with 377 respect to detecting community membership.Furthermore, 378 we find that hashtags are as informative as textual tweets in 379 inferring communities, while emojis are almost as informative.380 This is significant for two reasons.

381
First, in terms of informational entropy, emoji and hash-382 tags are more compact than text.Hence, they provide a 383 computationally-efficient way of determining community mem-384 bership.This is useful when dealing with large data-sets like 385 the one analyzed here.Additionally, and in contrast to text, 386 emoji and hashtags are freestanding expressions that can be 387 interpreted even before considering syntactic complexities like 388 word order or sentence structure.Based on nothing more than 389 conditional probability analyses (Figure 3) and a co-occurrence 390 matrix of emoji and hashtags (Figure 4), we were able to learn 391 a great deal about the various communities involved in the 392 Black Lives Matter discourse.

393
In addition to capturing the most obvious slogans (e.g., 394 #BlackLivesMatter) and counter-slogans (e.g., #AllLives-395 and trans people of colour are conveyed through the use of 397 emoji and hashtags.This result is significant in that it re-398 flects the movement's emphasis on giving voice to historically-399 marginalised victims of oppression (43)(44)(45).Consonant with 400 previous research, we are also able to confirm that the right-  3) suggests that this form of emoji use is not a sustainable 453 strategy.To the contrary, we find that emoji use is by and 454 large negatively correlated with retweet count and in effect 455 diminishes attention via engagement.
Finally, H4 proposed that emoji and hashtags are primarily affiliative gestures that call attention to individuals as bona fide members of a given group.In light of the evidence, this strikes us as the most plausible response to RQ2.For instance, Table 1, Figure 3, and Figure 4 all suggest that emoji are reliable markers of group affiliation.Hence, it stands to reason that they are also used as such.Further evidence in support of this conclusion can be inferred from the results of our retweet analysis (see Table 3 and Table 4).Recall here that, even though using emoji and hashtags in a given tweet generally decreases the amount of attention awarded to that tweet, doing so simultaneously increases the prospect of receiving more attention for future tweets.
Accordingly, it stands to reason that emoji and hashtags can be interpreted as affiliative gestures that impose an initial cost on signaling one's commitment to the group.Indeed, as as Figure SI1 shows, there is a substantial overall retweet penalty for both emoji and hashtags, with a correlation of −0.73 and −0.66, respectively between number of items and mean retweets.However, signaling does so with the payoff of increasing one's standing and following within that group.If this is right, our results suggest an interesting tension between H3 and H4.On the one hand, using emoji and hashtags signals commitment to one's group and increases the prospect of receiving more attention from that group at some future moment.At the same time, it decreases the amount of attention awarded to one's present tweets.
In sum, we started by noting that hashtags can be understood as indexing mechanisms that systematically codify certain topics under a common descriptor (23) and thus potentially connect people who are interested in those topics (10).Hence, we expected, and have now established, that hashtags are a strong marker of group affiliation.This is especially true for hashtags used by the conspiratorial wing of the Reactionary community (e.g., #qanon, #wwg1wga ) and the K-Pop wing of the Booster community (e.g., #matchthemillion, #matchamillion).We note that the use of hashtags to signal affiliation is not entirely foolproof.For instance, our analyses confirm that the Boosters at one point briefly dragooned various Reactionary hashtags such as #bluelivesmatter and #whitelivesmatter.However, trolling is not likely to be a sustainable strategy in the long term, and we would expect that the attention economy would quickly discard such efforts (as in fact happened in this case).
At the same time, we are surprised to see that this indexing function does not drive engagement.Contrary to previous research (49), we conclude that hashtags are negatively correlated with retweet count and serve a primarily affiliative function-at least as they are used in connection with the Black Lives Matter movement.
With respect to emoji, our expectations are again only partially confirmed.On the one hand, emoji such as the raised fist (� � �,� � �,� � �,� � �,� � �,� � �) signal in-group membership and community alliances among Activists, Progressives, and Boosters.Likewise, the American flag (� � � �) and various police-related emoji (� � � � � � � � � � � � � �) are all clear makers of belonging to the Reactionary community.Conversely, we find that interlocutory gestures like pointing fingers (�) and exclamation marks (�) do not succeed at drawing and directing attention to specific tweets.Hence, emoji too can be understood as principally performing a kind of affiliative function.Together, these results suggest D R A F T that emoji and hashtags play a complex role in the attention economy, operating at the level of both individual tweets and their authors.Additionally, the use of emoji and hashtags in bios-as opposed to tweets-remains under-studied.One hypothesis prompted by the current research is that here too, emoji and hashtags would play an affiliative role.It would also be illuminating to test whether the affiliative use of emoji and hashtags generalizes across other topics (or is constrained to the Black Lives Matter movement), and to examine discourse on other platforms to test whether this use is confined to Twitter.
These directions for future research reflect some of the limitations of the current study, which does not cover the full multi-year history of the Black Lives Matter movement on Twitter, let alone on all platforms.
As protests and counter-protests continue to migrate to the digital space, the need to understand the use of emoji and hashtags in online activism becomes increasingly important.
The current paper responds to this need and provides novel insights into the use of emoji and hashtags in online activism that we hope will be useful for further research.

Fig. 2 .
Fig. 2. Daily word-count sums of tweets associated with different communities.The vertical line indicates 25 May 2020, the day on which George Floyd was murdered.Date range: 17 January 2020 to July 23 2020.

232Distinctiveness.Fig. 3 .
Fig. 3. Conditional probability of community membership given emoji (left) and hashtags (right) for top 15 most popular in each community.

Fig. 4 .
Fig. 4. Co-occurrence network of the emoji and hashtags used by by each cluster.Node size = degree.Text and edge color = modularity class.

•
EH-tweets by EH-users have fewer retweets than either non-EH tweets by EH users or by non-EH users (e.g., Activists who sometimes use emoji have their emoji-using tweets retweeted on average 55.39 times, compared to 74.76 times for non-emoji using Activists); • where EH-users' non-EH tweets are more frequently retweeted than either (e.g., the non-emoji tweets by Activists who used emoji in other tweets are retweeted far more than either; by 105.6 times).Type II patterns: like Type I, save that EH-tweets by EHusers are roughly equivalent or have a slight advantage over non-EH users' tweets.

401
wing backlash to the Black Lives Matter movement is spear-402 headed by loosely-related, racist, and conspiratorially-minded 403 conservative partisans who come together in support of the 404 police and former president Donald Trump (42, 46).More 405 surprisingly, Figure 4 accurately captured the link-sharing and 406 fundraising efforts of a small but loud Booster contingent, as 407 well as the political ambitions of the Activists' Progressive 408 allies.409 Next, consider the use of interlocutory gestures and skin-410 tone modifiers.With respect to interlocutory gestures, we note the differential usage of 'raised-fist' (� � �,� � �,� � �,� � �,� � �,� � �) this hypothesis holds to a certain extent in each of the four 451 communities, the prevalence of Type I and II patterns (Table 452

Future
research could, amongst other things, examine the ongoing activities of the communities studied in this paper.Examples include the November 2020 US Presidential Election; the 6 January 2021 insurrection by Reactionaries; and Twitter's suspension of Donald Trump and purge of QAnon accounts.

ACKNOWLEDGMENTS.
This publication was made possible through the support of grants from the John Templeton Foundation (61378) and the Australian Research Council (DP190101507).The opinions expressed in this publication are those of the authors and do not necessarily reflect the views of the John Templeton Foundation or the Australian Research Council.The authors would like to acknowledge XXX, YYY, and ZZZ for their helpful comments in improving this publication.

Table 1 . Classification task results
these results confirm that the various 223 communities involved in the Black Lives Matter discourse use 224 emoji and hashtags in distinctive ways.What is more, emer-225 gent patterns of emoji and hashtag use correspond with pat-226 terns of retweet behaviour (Figure1), indicating that retweet 227 engagement is associated with the use of emoji and hashtags.228 In addition to this, we find that hashtags are a particularly 229 powerful marker of community membership and that, taken 230 together, emoji and hashtags are roughly as informative as 231 text in this regard.
emoji-obtained an accuracy greater than 0.5.More gener-215 ally, we note that deep learning techniques marginally outper-216 formed traditional approaches.More specifically, we find that 217 the best-performing classifiers were GRU and LSTM neural 218 architectures which take ordering into account, suggesting that 219 the order in which emoji are presented in a given tweet makes 220 a difference, perhaps because they occur in decreasing order 221 of priority for the user.222Withrespect to RQ1,

Table 2 . Descriptive statistics for emoji and hashtag use
more use of the link emoji (� �, associated with links to websites for petitions and donations).As outlined earlier, a substantial chunk of these users are associated with the K-Pop band BTS, which activated its followers by making a sizeable donation to the Black Lives Matter foundation.In addition to the link emoji, this affiliation is reflected through the purple heart emoji (�, a symbol of BTS) and a range of hashtags referencing BTS's million-dollar donation and the fans' efforts to match this donation.Usage Statistics.Descriptive statistics of emoji/hashtag (EH)