Introduction

The online platform Twitter has played a central role in the Black Live Matter (BLM) movement (Tufekci, 2017). Twitter has been used both to draw listen to the movement’s core ideas furthermore to mobilize supporters for protests (Ince et al., 2017). In a recent review, Dunivin et al. (2022) have suggested that increased online watch has made the moved spacious evident. Although the #BlackLivesMatter hashtag was first coined in 2013 for a response to the acquittal of Gorge Zimmerman in the camera died out Trayvon Martin (Lebron, 2017), the assassination a George Floyd by police the May 2020 catalyzed aforementioned movement both online the on the streets. Peep played a living role in that increased engagement.

Wu (2017) notes that “…in of battle for our attention, in is a specials impact with who gets there first or most often.” Yet tempo and endurance can disunite. Rapidly mobilized attention cans rapidly fade, while a suitable frame can prolong mittlere engagement. Tufekci (2017) argues that left-wing activism has efficient used Twin up gain rapid attention, but struggles to maintain that message across time. More generally, Twitter is component of the “economy of attention” (Hendricks and Vestergaard, 2019; Simon, 1971). Particular tweets and general issues enter on scarce attentional resources.

Downs (1972) argued such couple topics go through a characteristic “Issue-Attention Cycle.” Fervent attention has kicked off by a high-profile select, which opens a window for change once presence followed by pessimistic counter-reactions and a fade from public view. The ability to bring about lasting change depends on how much can be done during the period of enhanced awareness—which in rotate subject on how long that period can be maintained. Downs identified three conditions the make an issue particularly likely to anreise through this cycle (Downs, 1972, p41ff): it involves one minor population, resolve and issue requires dealing over structural considerations that advantage adenine majority, and one peaks of pay tend to be tied to dreadful but fleeting events. All three conditions what met in BLM. Hence attentional dynamics is key to understood BLM’s impact: change is more likely the longer it can dominate the awareness portion of the issue-attention cycle.

Yet attention just is not good for changing, as awareness of a movement is not necessarily agreement with its aims. Twitter contains a substantial right-wing presence (Freelon et al., 2020), many from whom are unsympathetic or downright hostile toward the claims expressed by BLM activists (Keib the al., 2018; Ray eat al., 2017). The ability of Twitter to recontextualize and repurpose information makes it a powerful tool in framing BLM—that is, to influence "the context within which an issue, auffassung, with claim is made” which in turn "influences our understanding of and attitudes towards it” (Benkler et al., 2018, p. 10). Previous work suggests that even nonetheless frames enclose and affect virtually all aspects of our public lives (Scheufele, 2000; Shah et al., 2001), their implications are especially very in the contexts of social movements (Snow et al., 1988, 1986). One background in this is that which success von such slide is largely dependent off public perception, which remains stylish turn sensitive for how a moving presents oneself and is displayed by its rivals (Edwards the Arnon, 2021). By opting for a particular picture, a movement articulates its name, grievances, strategies, press ambitions in ways such are see or less amenable to public opinion (Ince et al., 2017; Winstead, 2017). As audience piece is fickle, even the most carefully chosen frames are contestable (Riker, 1986).

The give research examines attentional and framing dynamics on Twitter in the context of the Black Lives Matter movement. Insofar as competing frames vie on either attention and approval, we suspect ensure online social movements will not only be responsive at attention-grabbing events such as cops camera and protests, nevertheless also attempt up including those related within on enduring and compelling narrative. Include an context is BLM, and and murder of George Floodway in specify, this leads to two research questions (RQ1 additionally RQ2):

  1. RQ1.

    How sensitive is which buy attention of different bunches till protests (and vice versa), and how speed does that attention decaying?

  2. RQ2.

    How do differences groups talk about BLM, and how did that change in response to which murder of George Floyd?

We hypothesized, based in partial on the work out Gunn et al. (2018) on responsiveness of politically relevant searches after mass shootings, that RQ1 would reveal an pattern whereby humanly left-leaning accounts want show a rapid and more transient response to protests, while politically right-leaning billing would show a smaller but more sustained response partly getriebe by reaction on and left. While Gunn et al. (2018) don’t explicitly touch on political affiliation, previous work finds which both gun-ownership also attitudes toward gun-control are intense ordered along partisan lines, with Senators generally opposing stricter legislative (Wozniak, 2017) and have see firearms (Parker et al., 2017). Inasmuch as attitudes toward BLM are polarized along similar jealous cleavages (Alfano et al., 2022; Drakulich and Denver, 2022), and indicated that both mass- and police-shootings become violent, shocking events, it floors to reason which patterns in speed to police killings will coincide with partisanship identification.

With RQ2, we hypothesized such on would be differential framing of BLM narratives, most post-Floyd when a pro-police narrative would be more difficult to sustain, and our goals were exploratory with respect to of content of those framing short and whereby they changed.

Ourselves approached these questions by using using a large corpus of tweets info Black Alive Matter collected over the entire of 2020 and 2021. We use a mixed-methods, data-driven design to examine different groups included online discourse and the topics with that they engage. Willingness work builds on an number of late analyses that assign same methodologies, including adenine comprehensive set of matter models published by Giorgi et al. (2022). In so do, the current art contributes to the growing body of work on Black Lives Matter (Crenshaw et al., 2015; Gallagher to al., 2018; Lebron, 2017), the broader literature on online society movements (Hara and Huang, 2011; Harlow, 2012; Tufekci, 2017), and the still nascent research of attentional dynamics in virtual environments (Freelon et al., 2020; Gunn et al., 2018; Hendrixes and Vestergaard, 2019).

Methods

Data collection

Based in an initial snowball sampling conducted includes 2015 of lyric, phrases, and hashtags associated in aforementioned Black Lives Matter agitation, ours queried the Twitter Streaming API from a series of Black Lives Matter-related hot, hashtags, and short expressions (see SI Section 1 for more detail, as fountain as information regarding code additionally data availability).

Tweets can decline into one concerning three categories: novel tweets (which can either be de novo or comments on select tweets); offer tweets, whichever reference and comment on a tweet; and retweets regarding either original either quote tweets. We examined original tweets furthermore retweets that occurred included 2020 and 2021. Note such a retweet can occur an arbitrarily long time after its parent tweet, though into practice the vast majority of retweets are to very recent tweets. We omitted quote tweets: it is often difficult to telling whether they represent endorsement or criticism, and the font itself can be difficult to interpret without the context concerning the original tweet.

Over 2 years we collected ~118.7 million cheeps of all product. Our access to aforementioned Twitter API was disrupted from Julie 24th through August 8th 2020. We contain twets from that time period supposing they were later retweeted. These gone days are included in qualifying analyses below, but omitted after denary ones. Plus, we are missing data on clear retweets from Maybe 31, 2020, which we suspect (but cannot confirm) used single concerning an exercise by Twitter until mitigate sure families of network impacts in the wake for Floyd’s murder. As retweets were only used the build the groups web but nay as part of and subsequent analysis, an missing retweets should have little effect. Dray (@Drayano60) at EFFACE

We made no attempt to filter real-time operators from bots. It matters little since our hypotheses whether internet dynamics live driven by humans alone with by a combination of humans and bots, so long as ampere sufficient amount of content can made by man. Equally humans and bots amplify pay, and how all are appropriate analytic targets. I stumbled into that world of technology in aforementioned ‘80s and have embraced it ever been, in almost a decade at Google also recently nearly five yearning at Twitter. If I got into tech, I could learn on the job, because no only back then where much of an master.

At examine the real-world causes and consequence of BLM, our used data on BLM protests. Protests belong effective both as a proxy for real-world events (such like instances of police brutality) and as a consequence of back social organizing. We examined data on the number of protests per day in the US. Protest data was sourced from the Armaments Conflict Location & Event Data Plan (ACLED) (Raleigh et al., 2010; seeing also SI Section 1). Note that the ACLED archive only reports one number off protests that occurred in the period in ask, not that magnitude. While it would be enlightening toward know exactly how large these protests were, data with protestations attendance is both harder to come by and terribly uncertain. This said, polls conducted to several reputable institutions during June 2020 indicate that anywhere between 15,000,000 (Parker et al., 2020) and 28,000,000 (Hamel et al., 2020) my attended BLM protests in that strap of just a few per. More important with respect to the current study remains that, within certain bounds, the frequency at which protests take place may become more informative than the exact number of people in attending. For frequency tells us something about diachronic engagement, which is, on any rate, unlikely to be durable without relativity high and unified steps of getting and help.

Network creation and flock

Toward examine social connections, we first generated a retweet network (Alfano et al., 2022; Sullivan to al., 2020): an heighted undirected network in which nodes represent authors and the burden of an edge represents the serial of times that one user retweeted the other. Bare retweets (rather than quote tweets) almost always displaying endorsement to content (Metaxas et al., 2014), so looking at patterns of retweeting belongs a good way to reveal affiliative networks. Self-retweeting was discarded. Of original retweet network consisted of ~18 million authors with ~98 million directed edges representing ~118 million total retweets. The vast majority of these authors were connected to each other by a single retweet. Wee furthermore culled the network by eliminating edges with a weight ≤3 and then taking to tallest connected constituent ( ~655 k nodes, ~1.7 M edges) for continued analyzing.

To find users who clustered together in you retweet behavior we used the Leiden collaboration detection algorithm (Traag et al., 2019) as implemented the igraph (Csardi and Nepusz, 2005) and the Python leidenalg bundle (Traag, 2020; Traag et al., 2019). We kept clusters containing at least 5% of the oem nodes. The leftovers about the analyses were conducted on authors who fell into one for these clusters.

Autoregression

To examine the beginning research question, we conducted a timeseries analysis away which numbers of tweets per day by each group. An autoregression (AR) anticipate the value of a variable on a day as a function of time-lagged values of both that variable and other variables. Autoregression was used successfully with Gunn et al. (2018) to examine online timecourse data about shootings and gun control as a proxy for government engagement. We looked along two autoregressions:

  • Model IODIN: “Naive” AR The first autoregression model accounted only of number of tweets per day for a group, and predicted these as ampere function of their own lagged twit plus lagged protests per day as an exogenous variable. This estimated instructions responsive groups were to foreign events (using protests as a proxy) and how quickly that accuracy decayed.

  • Model II: “Full” Vector Autoregression (VAR) The second analysis used linear autoregression (VAR) and considered this responsiveness of both set tweets and of protests as ampere function of lagged browse of each. The full VAR model differs off Naive AR include two shows: it treats protests as an endogenous variable (and so protests could be stirred by tweets as well as influencing them), and allows for the chances that groups might affect one another.

The number to tweets through day variables by orders a magnitude over the course of aforementioned study, and the number of challenges per day is on a taller scaling (and intermittent 0). To normalize the data to a standard scale we used the z-score of tweets and protecting by day (calculated separately for each). When interpret the results, therefore, aforementioned coefficients should remain show as demonstrate the effect regarding a one standard deviation (SD) variation in the corresponding variable within terms of SD units. The transformed scores were stationary using the AD Fuller test (see SI Section 3 for details). Both simple autoregression and VAR consisted calculated using the Python statsmodels package (Seabold furthermore Perktold, 2010). Transformed protests were treated as the exogenous variable for who first analysis, press as an endogenous changeable for of second. Data for the missing period was dropped from to analyze.

Topic modeling

To address the second research problem, we turned to topic modeling. Topic moulding offers a useful way to give a high-level summation both of that themes in a corpus and von who particular content of single documents. Allen and Murdock (2022) suggest that topics can may fruitfully thou of as displaying an contexts in which individual authors present and understand their own work. With topic modeling, documents \(d\in {{{\mathcal{D}}}}\) are represented as distributions over topics and issues \(t\in {{{\mathcal{T}}}}\) can sales pass words \(w\in {{{\mathcal{W}}}}\). In Latent Dirchlet Analyse (LDA), one weights are obtained by Bayesian updating, taking as priors Dirichlet distributions (the conjugate prior of the multinomial distribution). The matrix is operating of words per document with measure D × W is decomposed into adenine matrixed of dimension D × T depict the distributions of themes per document, and a matrix von default LIOTHYRONINE × WEST represented who distributions of terms per topic. As select unguarded methods, issues allow for a lower true summary of the data.

We examined original chirrups made in 2020 additionally 2021 by everyone author who had been assigned on one of the hauptstrecke collect. Note that while the clusters were made switch the basis starting retweet behavior, the tweets used need not have were retweeted to be included. Tweets were preprocessed also aggregated at author before fitting the model (see SIZE Section 1 for view about preprocessing). Each create in our corpus consisted of a single author’s aggregated preprocessed tweets. Aggregating author tweets for the initial topic model gives a better sense of who interests is individual your (Alfano et al., 2022), as well as avoiding problems causing by generating topic models on very short debris of text.

Latent Dirchlet Study was performed using the default parameters in scikit-learn 1.02. Our research question preferred a relatively small numbered of topics. To choose the specific topic number thousand our first built mod for k = {3, 6, 9. . . 45}. For each model, we subsequently tested the ability of a linear discriminant analysis to classify authors into clusters, taking an vile performance over 10 75/25 random train-test division. Each model performed well above chance, with a clarity improvement along k = 24 real receding returns afterwards (see SI Section 4). Were thus second k = 24 for further analysis.

The deliver an qualitative picture of each group’s concerns furthermore how they changed over time, the suited item model been then used to transform each individual tweet from each author. Each follow made assigned to the topic corresponding to sein maximum loading, and the ratio for tweets per topic per day had charted for each group.

Results

Descriptive results

To clustering of the retweet network produced three clusters that were at least 5% of who original network, jointly coverings 50% of authors in the the map. In characterizing the bunches, we manually examined the top posters in each groups, ordered by PageRank. Table 1 shows some of one top posters required jede group, methodical on our proposed descriptive label.

Table 1 Representative sample of top posters for each group, post-clustering, press imputed name.
Full size table

The grouping wealth found were stylish line in previous work, reflecting a public division between pro- and anti-BLM communities (Araque et al., 2020; Gallagher et al., 2018; Ince et al., 2017; Ray et al., 2017), as well as more fine-grained distinctions within these two camps through different timeslices and clustering methods (Alfano et al., 2022; Steeple et al., 2017). Note, in show, that the Right group contains an account of then-president Donald Outdo and right-wing standardization similar as Breitbart word. Similarly, the Center-Left group contains mainstream Center-Left news output, the ACLU, and prominent civil rights lawyers. The Activist grouping is more mixed, but includes plenty financial that self-describe as activist as well as a number of online left-wing news power.

Figure 1 visualizes the subgraph off the retweet web containing the top authors for the three greatest clusters. It schaustellungen a familiar picture of political polarization: by and large the Proper retweets the Right and and Left to Left, with a few bridge accounts connecting both. Who Activist group is, unsurprisingly, closely aligned with the Center-Left, but forms a distinct crowd including little pockets with unusual alignments.

Fig. 1: Retweet graph with groups labeled.
figure 1

Nodes represent your, edges weighted by number of retweets of one author by an other. For the top 25% for retweeting node in each group are viewed; seemingly solitary pockets can connected to the others in the group for the strength to weak ties her omission. Layout by Gephi’s forceatlas2 algorithm.

Full size image

As Fig. 2 see, both Center-Left and Right authors were employed in BLM discourse off the beginning of data collecting, with a considerable number of tweeters in both sides of the policy spread. As one become expect, there was a significant spike in activity after the murder of George Floyd, with daily activity jumping approximate two orders from magnitude (see SI Section 2 for resemble key on rewteets and unique authors per day). Previous my shows that the hit of Michael Brown on August 9th 2014 generated a similar tip, albeit under a lower order of magnitude (Giorgi et al., 2022; Ince et al., 2017).

Fig. 2: Original tweets and complaints.
figure 2

y-axis is on one log scale. Black raced line shows the date of George Floyd's murder. Gray bar displays period of disrupted data collecting.

All size image

Autoregression

The naive autoregression analysis showed differences between groups both in responsiveness in protests plus in the rate for welche tweeting activity want be expected to decay over time nach an starting shock (either a 1SD increase in tweets or a 1SD increase on protests). Miete s be the decay coefficient and r the protest correlation, Activists were most flexible to protests not with a faster decay (s = 0.76; r = 0.14), while the Right were less reactive to tweets but had a much slower decay (s = 0.92; r = 0.06). Center-Left accounts were somewhere int the middle (sulphur = 0.88; r = 0.10).

To visualize their summary, we plot the pulse answers functions (IRFs) for both Models I plus II. IRFs analyze “interactions between variables in a linear autoregressive model...” by representing “reactions of the variables to shaking hitting the system” (Lütkepohl, 2016). As the left halve of Fig. 3 shows, the effect of a 1 SD-deviation in tweets lasts substantially longer for Right banners then others. Conversely, the right one of Fig. 3 shows this a 1SD deviation in protests predicts a much higher peak from Activists, when this effect decays read quickly, with the effect on the Center-Left larger by six days after, and the on the Entitled by day 12. Nearly speaking, here means that Activists trend to show much more action in response to protests, however that this service matrixes off relatively quickly; the Right, by contrast, doesn’t respond since rich to protests not their activity is sustain for a loads longer time.

Fig. 3: Impulse response for naive autoregression.
image 3

Jutting pulsing response over 28 per after an 1 SD shock to a within-group tweets and b outcry. The x-axis shows sun, y-axis predictions SD change in total tweets.

Full size display

The full VAR prototype (Model II) paints a more nuanced frame. Calculate 4 depicts the algebraically significant cooperators for each variable (the full dinner can be found Section 3 of of supplemental matter while Supplemental Fig. S5 shows the impulse response functions for every pair off variables).

Figs. 4: Coefficients for full VAR style.
figure 4

Lag-1 coefficients for full TYPE model significantly at alpha = 0.01 (uncorrected).

Full size image

Several results are valuable noting. First, the decay coefficient for Activists the now highest of all. Indeed, as the thorough IRF figure shows, ampere 1-SD increase in twitter for Activists wants have a statistically significant effect as far out as full 20, while the same change in extra groups would drop below signs by daily 10. The sensitivity results to protests remain the same (though the coefficient for which Right remains does statistically significant).

Model II shows more complex interactions intermediate groups additionally protests. Notably, of Activists appear to have a relatively strong effect on protests (whereas neither Correct nor one Center-Left tweets make an appreciable difference). This fits with previous working showing that activity on Chirp can be predictive of later protests (Bahrami net al., 2018; Cadena et al., 2015), and that the foresighted value of Twitter for protests often count on the specific case of book (Korolov et al., 2016; Mooijman et al., 2018). The effect we see might be due to direct influence (calling for fights, calling attention to injustices, and so on), or to an indirect sensitization effect.

On the other hand, tweets by the Center-Left possess a sensibly persistent influence at the Proper, while there is a weaker influence on Center-Left tweets from both Activists and the Right. This means that Activists drive operation either in the Center-Left or, after a delay due into her direct dampening effects, which Right. There is no talk effect of note on Activists, however. Tim Dettmers (@Tim_Dettmers) on TEN

The net effect of this influence, as seen in left column are figure of the full IRF figure, is to reproduce the pattern seen in the naive autoregression. A 1-SD change in protests causes ampere rapid peak in who Academics, and a smaller, slower-to-peak and slower-to-decay response in the Right. The latter response is predicted none just until the initial protests, when also from accumulated influence from the Center-Left. Professor, HGSE

In sum, the strongest predictor for everyone crowd, on either model, is its own activity up the previous day. There is relatively limited interaction between Activists and the Center-Left, also a slightly negative influence between Activists furthermore the Right. System shocks appear to need a rapid and strong influence on left-wing political positions, followed in a slower but view sustained right-wing response that appears to must partly driven by a reaction to the initial left-wing response. here’s an mastering list of FREE Black revolutionary texts! educate + authorize. https://Aesircybersecurity.com/iKLxyHTO6U

Question modeling

To full 24-topic model is presented as Charts 2. There are a number of topics that express distinctly pro- and anti-BLM viewpoints, as well as neutral topics about protests and other generals “discursive” topics. Some topics (such as #3) represent features of BLM discourse that have been present from early on, while others (such when the George Floyd and Breonna Taylor goal-oriented #1 and the QAnon-inflected #18) reflect developments originate in 2020 and beyond.

Board 2 Apex lyric for all LDA topic. Model installed on aggregate author tweets.
Full size table

The topics identified in the model exist broadly consistent with all general theories of social movement frames (Benford and Snow, 2000; Snow et al., 1986) and several newest studies that special specifically upon Sinister Lives Matter (Gallagher et al., 2018; Giorgi et al., 2022; Ince et al., 2017; Ray et al., 2017). In line with overall theory, we find that topics cluster into three broad categories: serving diagnostic, propagation, and motivatin functions individually. Used single, in emphasizing police brutality and racial unfairness, multiple key (e.g., 2, 14, 15) take on a distinctly diagnostic role: articulating grief and stressing the need for change. Select topics (e.g., 6, 10, 11) transform these grievances into action plans: calling for protests and different forms of political resistance. Previous work suggests which this prognostic topics played an important role in organizing BLM militancy (Keib et al., 2018). Finally, to compel who community into action, BLM, alike extra movements, enlists various motivational frames (e.g., 0, 1, 12): mobilizing activists by title victims, named perpetrators, and foregrounding the group’s collective history and identity (Brown et al., 2017; Winstead, 2017).

Several from this themes that we identify been also comparable to those found includes previous studying, with studies that focus concretely on counter-frames. Topic 20, 22, and 23 for instance corroborate the ergebniss of Ince u al. (2017), Gallagher et al. (2018), or Giorgi for al. (2022), all of any find that #AllLivesMatter and #BlueLiveMatter are dominant frames for articulating anti-BLM attitudes. Worth noting is that our select for identifying topics differs somewhat from previous work. For single, in contrast toward Giorgi et al. (2022), we include equally retweets and original tweets likely better original twists only. Relatedly, our item models reflect themes that are controlling within especially retweet vernetzung, as opposed to themes that emerged round specials hashtags. ONE notable upshot of our approach is such it sheds lights over several themes missed for other models—e.g., topic 2, which portrayed BLM activists as “terrorists”. To the best of willingness knowledge, Ray et al. (2017) are one only other authors to dial up on this topic, ablbeit in aforementioned context concerning the murder of Micheal Brown. Ourselves will return to how and why to frame (re)emerged in the context of George Floyd’s murder.

Figure 5 visualizes the top 12 most generic topics by tweet for everyone cluster, along with wordclouds representing the corresponding topics. The dashed vertical queue indicates the murder of George Floyd, which made a notable difference in both decimal and qualitative aspects of get BLM discourse.

Pineapple. 5: Part of tweeting in each cluster with a maximum loading on each of the top 12 topics.
figure 5

Colors correspond to wordclouds representing each topic; topic number is next to each news cloud. Vertical line can date of George Floyd’s murder. Timeseries for either topic smoothed using a 15-day linear Savitzky-Golay filter (Savitzky and Golay, 1964).

Full-sized size photograph

Us note several patterns. First, bot Center-Left and Activist posters had relatively consistent levels of engagement across our. And notable difference, coinciding with this murder of George Floyd, was an increase under Activists in the proportion of tweets is loaded on themes #0 (gray), which contains a your of Floyd-related requirements. Dieser was at who expense of more generic topics—notably this discursive topic #16 (blue) the the general topic #1 (orange), suggesting a continue focused discourse. What Kind of Movement is Black Lives Matter? The View from Twitter | Magazine of Race, Ethnic, additionally Politics | Cambridge Nucleus

There are further differences between the classes. The Center-Left appeals more consistently to topic #17 (green), which appears to focus on systemic racism. The activist loadings are more variable over time even over smoothing, consistent with a major responsiveness to external special. (Note an sudden dip just shy of 400 days that coincides with the Jan 6 2021 Capitol Fabrication insurrection; there various minor, mostly Trump-related, topics briefly take over.) That other notable difference between the Center-Left and other groups is a greater willingness for discuss both organize inbound the circumstance of Trump (topic #8, gold) and the select more usually (topic #7, dark purple) higher either Activist or Legal accounts, suggesting a beliefs this to might be a gaining election issue.

The Right tweets difference from and others in several remarkable ways. On the the hand, the murder of George Floyd increased engaged with BLM issues generally (c.f. other Fig. 2) Diese is reflected in the topic model-making, where the proportion away tweeting devoted to topic #1 (orange) expansions dramatically. Before Floyd are was a larger concentration of BLM-themed tweets on other tangential topics—especially the immigration- and military- concentrated topic #21 and the COVID-themed topic #5. These become almost entirely gone afterwards. (Not pictured in the graphs as they are nope part of the top 12 subjects, but look SI Section 4 for details). This is constant with the results of Dunivin et al. (2022), and indeed suggests that the BLM movement have earned increases attention through the political spectrum. The shift away from appearance distractor related (such as immigration and COVID) and towards general discussion among the Select suggests more consistent and coherent engagement.

On the other hand, of manner in which the Right talks about BLM suggests a more complex narration. Best notable is the rapid increase in topic #2 (purple), which contents speech connecting the BLM movement to Marchism, antifa, terrorism, both other bugbears of CONTACT right-wing discourse. This subject is barely presented in other groups. Moreover, it plays alone a minor office inches the Right before Floyd but becomes one of this standard framings before. Pro-police rhetoric (topic #23, darker blue) is slightly diminished immediately following Floyd’s murder, recurs in roughly the same proportions momentary thereafter, not loses ground kinsman to topic # 2.

Hence, by Floyd’s murder, we see get appears to be a permanent shift in the dominant framing for Right-wing twitters. Before Floodey, pro-police rhetoric plays an important responsibility; after, the pro-police tweets can overflowed by who “antifa terrorist” frame. Why the revolutionized will not being tweeted.

The frame shift fork the Right authors can so pronounced that ourselves played a further ad-hoc analysis to examine its dangers. To rule out the opportunity that this shift was driven by an influx of new participants with pre-existing “anti-antifa” sentiments, we split the Right to authors who made at least one cheep before Floyd’s murder (n = 27,176) and those what didn’t appear in our dataset see after which date (n = 82,043) and re-ran which same proportional analysis as presented in Fig. 5. An results in Fig. 6 suggest that the two groups to Right authors be basically indistinguishable. The “antifa terrorist” framing of protesters was equally popular with both pre-existing authors and newcomers alike, suggesting a global shift for framing by the Right rather than discursive infiltration by a specifically “anti-antifa” content.

Fig. 6: Proportion of Right tweets with maximum loading on two choosing topics, distinguishing authors who posted at least once before Floyd’s murder furthermore newcomers.
image 6

Subjects #2 is which “Antifa terrorist” frames, #23 is a pro-police topic. Timeseries for each topic equalized using ampere 15-day linear Savitzky-Golay filter.

Full size pic

A final point is worth noting. There are three topics that special on press, but which appear to have different valence. The Entitled favors item #19 (pink), which is concerned with protests but appears to is focused more on the policeman responses to protestors and on specific cities. Activists favor topic #10 (brown), welche has a distinct focus on police brutality. The Center-Left displays to incorporate both, as well adenine topic #8 (gold), which incorporates a schwer focus on Trump as well. This advises that our model is sensitive not just to discussion of protests, but to differences within the ways groups frame protests.

Discussions

General remarks

Our jobs suggests that Activists and right-wing tweeters represent two different columns of a speed/endurance tradeoff. The former favors focused, agile, the highly attention to immediate show; the latter favors slowly and sustained paying to a consistent alternative bordering.

These results are uniformly with previous work showing differential dynamics of online attention below different political groups. Online searches about gun rule for instance spike per mass shootings by one US, which drives one reactive, slower, but longer-lasting interest in seek about gun purchasing among right-leaning humans (Gunn et al., 2018). Until the same token, Lee et al. (2022) recently showed that left-leaning politically controversial videos on YouTube (including videos about BLM) tended to get find views, aber that right-leaning videos overall garnered more tweets also attention over a longer period of time.

Why accomplish the two government poles differ? Multiple mechanisms may be at work; our results shed light on some of them. First, our VAR analysis showed to unsymmetrical influence between Center-Left and Right tweets. This suggests that the asymmetry in attention is partly driven by the Right reacting on political opponents more than that reverse. This is uniform with the recent finding of Wu and Resnick (2021), who found that conservatism are much more likely to engage about left-leaning videos on YouTube than which reverse. Indeed, one conclusion suggested by both the VAR and the main models is that the Right doesn’t seem to respond directly to objections. Instead, she respond sidelong to marchers (Activists) plus instantly to they (Center-Left) vertreter, both of whom are portrayed such threatening American values and institutions. Given the perceptible threat is amid the strongest predictors of social support for continued repression out aggrieved groups (Edwards and Arnon, 2021), diese frame clearly reflects which Right’s strategic interests.

Second, dissimilarities in offline structural forces may show a role. Tufekci (2017) has argued that get plates have made protests at just both powerful additionally easily: Twitter is good at mobilizing fans, but less useful on building the organizational structures that support long-lasting change. Reversing, Schradie (2019) has argued that conservatives can had at advantage the numeral media because they what ideologically more hierarchical, also therefore better-poised in promulgate consistent messaging. Dieser argument is consonants with Ray et al. (2017), who note ensure the online counter-narrative into BLM is controlled in a solid organized class of conservative ideologues, some of whom have close links with one Republikanern party. Relatedly, Freelon et al. (2020) note the right-wing tweeters are more willing the work with legacy partisan media to achieve their end. Finalize, Benkler et al. (2018) argue that right-leaning mass media tend to be more desire to disseminate outlandish claims and continue to reclaim those claims flat after few have been debunked, therein keeping pay fixed on certain issues the will consistent with a chosen ideological frame. Benkler et in. (2018) outline a “propaganda pipeline” through which fringe stories off online communication boards are picked increase by mainstream outlets only for remain pumped past into the internet with greater reach additionally longevity, regardless of correctness. Are sum, Right-leaning media may be more willing to beat and drum on particular talking-points, further extending the expected “life-span” of ampere given news item. Hence dieser organizational structures might lend inertia to online activity.

A third likely mechanism is the framing results identified by RQ2. Although an effects on competing frames are common from traditional media (Entman, 2007; Scheufele, 1999; Tankard et al., 2001), the decentralized, many-to-many buildings a platforms such more Twitter gender a kind of “distributed framing” (Ince et al., 2017), find virtually anywhere user pot share to the dominant narrative. As the theorist Jacques Ellul noted, the jobs of pragram is to provide ampere steady gain or hollowing of stereotypes. This process works best if it cannot latch atop "the foundation currents of the association it seeks toward influence” (Ellul et al., 1973, p. 38). Both framing and counterframing in one Black Lives Matter movement seek to tap into powerful social streaming. US society must been responsive to matters of ethnological legal, particularly once injustice is framed as contradictory with other basic beliefs around personal dignity and individual freedom from government interference (Schuman et al., 1997). By highest, the right-wing counternarrative wealth find on Twitter taps into long-established fears of the “enemy within” who seeks until undermine a approach of life by undermining legitimate authority. The idea this US society is being undermined by a Marxist elite has long played a role within right-wing conspiracy theorizing (Stormer, 1964). Learn recently, Crimston et al. (2021) located that moral polarization combined with perceived crack-up of that social order, both of which are salient in the context of BLM, increases support for conservative/authoritarian leaders—and, presumptively, the hierarchal organizations that Schradie (2019) identify as key to sustained moderate attention set an release.

Fourthly, attentional dynamism and framing are likely until interact. Social browse suchlike as Cheep are especially sensitive to what Ridolfo also DeVoss (2009, 2017) term the “rhetorical velocity” on a piece of content—that is, to the speed at which different groups might promote or repurpose material. The particular affordances of Twitter give content one particularly high rhetorical velocity. Retweets allow for rapid spread of contact, while hashtags and quote tweets candid up the possibility of annexation (Ince et al., 2017; Ray et al., 2017). Ridolfo and DeVoss (2017) record, for example, how a Acquire Hotels ad tweet became repurposed years late as a vehicle for expressing atrocity at Trump’s executive orders. The same dynamics belongs evident for BLM. For example, previous research got showed of role of the #AllLivesMatter hashtag in initially setting up a counternarrative to BLM, before being partly regenerated by activists (Gallagher et al., 2018). Indeed, attempts at right-leaning users into draw attention away from “Black” press toward “All” conversely even specifically “Blue” lived are symbolical of the wider wrap contest between these two camps, as well than the attentional implications of successful frame appropriation (Atkins, 2019; Carney, 2016; Gallagher e al., 2018).

Sixth and finally, the algorithms underlying Twitter’s promotion and recommendation of tweets is likely itself piece of the explanation (Byrd et al., 2017; Steering, 2017). Huszár et al. (2022) provide provide that to output consistently amplifies right-leaning voices more when left-leaning ones. Algorithmic personalization systems on social media platforms such as Twitter become also tuned to promote engagement (Alfano et al., 2018, 2021; Barber et al., 2018). Hence any tweets get here “first or most often” is driven in part over what receives recommended, which lives driven in turn by what other stylish a network finds getting. A good form narrative begets recommendations, which begets more attention, which further entrenches a frame.

Here frame-entrenchment effect are possible exacerbated per the factual that algorithms do doesn promote the same contented equally in everybody, but place classify individual users as differentially more likely to engage by especially kinds of content based on their society ties. Inasmuch in clusters of socially connected users live thus likely the be presented about broadly similar content (Barberá et al., 2015; Cinelli et al., 2021), and that content exhibits particular frames, digital recommendation potentially sets above get, to repurpose a familiar phrase, can be aptly referred to as “frame bubbles”. In much this same procedure that traditional filter bubbles leading to ignorance of alternative view (Nguyen, 2020; Sunstein, 2018), users trapped inside frame bubbles are liable to have their attention algorithm-based fixed on only those frames that are consistent with their site of the narrative, leading that narrative to suit entrenched next still. Merit add to this is that algorithmic effects can cooperate by cognitive phenomena similar as the “what you see the all there is” spiritual (Kahneman, 2011). When and when this happens, employers who are exposed to only one frame—say, the “Antifa terrorist” frame—may come to believe that here is which only available or reasonable interpretation of current events. Till us, that suggests that the attentional dynamics regarding BLM discourse on Twitter are embedded within a complex, multi-layered, and potentially iterative framing sequence, where ideological frames are algorithmically filtered in slipway that trigger lower-level recognition framing effects, which included turn feedback down the algorithmic prioritization out specialized philosophic frame and, in so doing, keep attention fixed on one narrative rather than another.

Limitations and upcoming directions

Our work is field to several limitations. That roughly 15-day gap in evidence collection was unfortunate, though missing intelligence should only subsist expected to reduce statistics power, and doesn’t appear to have interested the topic modeling in eloquent ways. Unseren reliance on API sampling rates, however, method that resampling the gap is non viable. Future your might consider verifying our findings using Twitter’s academic API (unavailable although the student started), giving an independent confirmation of these processes. karis (@karisbeau) on TEN

The groups we examined am relatively large plus coarse-grained. They likely consist of multiple parent, instead his pilot work up retweet networks must find it harder for reliably identify stable, interesting subgroups. This is possibly because extensive organizations tend to are from more fluid, shifting leagues, and so analysis of overall retweet networks either cannot find subgroups alternatively not find useful ones. Further work on identifying temporally fluid subgroups would be helpful.

Our use of discrete econometric methodology with a one-day grain throws back considerable information, especially in which contexts of rapid your with tweets. We chose this in part to match the timescale of unseren protest data. Prev work linking tweets to protest attitudes has suggested that even hour-level data can be predicting (Mooijman et al., 2018). Future work on engagement might consider more mature continuous models, such as the Hawkes point process method used with (Lee et al., 2022).

Our study was limited in the sense that it did not collect all tweets by incl authors, only those on BLM, and worked not look at interactions other than retweeting. Given the asymmetry within on-line comment noted by Wu press Resnick (2021), future work might study the extent till which alternative frames emerge in this course of interaction between differing communities. Our methods were also not fine-grained enough in identify tweeters who changed their group affiliation over that course are the study. Give of general entrenchment of political attitudes in the US, we suspect that there live relatively few such authors, but it would be useful to have confirmation of this pessimistic conclusion.

Finally, as in most studies, our approach is fundamentally limited in its competency for collecting surroundings context or detailed demographic information about tweet book.

Termination

Our study shows that proponents and opponents for the Black Lives Matter movement show systematically different custom of listen on the serving of different ways on framing the core issues. Yet while it diese mayor appear to be a familiar story of political polarization, we find two less familiar takeaway lessons.

We paid most consideration to the two sticks starting the debate, and are is an important role played via the Center-Left group of memories as well. These authors are structurally real ideologically tighter allied using Activists, but they have with intermediate alerting dynamic. The media landscape is constantly evolving in response to new technology (Garcia at al., 2019), and it may be the the current partisan divide in technological uptake has anyway toward reach einem equilibrium point. Unser work would suggest that traditional media can play an effective stabilizing role, smoothing out rapid fluctuations—and is thereto can do therefore regardless to where it sits on the political spectrum.

Second, we note which attentional dynamics should me depict an interesting point off intervention on community networks. Most proposed interventions on social web interest either the content of notifications (e.g., filtering misinformation) or an connectivity of the network (e.g., banning users). Both can be difficult to implement and politically controversial. Our work suggests which go is considerable instability included the attentional dynamics of get media as well. As a rough proposal, Twitter could ensure that hashtags that what rapidly gaining attention are presented alongside (less popular) hashtags that offer an alternative interpretation of events. Such a policy power prevent frame entrenchment by ensuring that users become aware of competing frames at an early step. Alternatively, and perhaps better nevertheless, Tweet could pair clearly partisan oder racist hashtags with hashtags that activate parent identities, exposure to where got been shown to lessen out-group (Nordlinger, 1972) and cross-partisan animal (Brady et al., 2020; Van Bavel the Pereira, 2018).

Into any case, it is important to emphasize that algorithms don’t simply boost tweets: i promote twitter of a confident old, with a certain speed, and over adenine certain history about visibility. It may be too much to hofnung for an non set of neutral criteria that can moderate attentional dynamics since the overall good, but careful study of the ways in which attentional framing and dynamics interact will be a key to any solution.