Attention and Counter-Framing in the Black Lives Matter Movement on Twitter

The paper investigates how attention and framing dynamics on Twitter shaped discourse around the Black Lives Matter (BLM) movement, especially after the murder of George Floyd in May 2020. Motivated by theories of the attention economy and framing in social movements, the authors pose two research questions: RQ1: How sensitive is the online attention of different groups to protests (and vice versa), and how quickly does that attention decay? RQ2: How do different groups talk about BLM, and how did that change in response to the murder of George Floyd? The authors hypothesize (for RQ1) that left-leaning accounts rapidly mobilize attention that decays quickly, while right-leaning accounts show smaller but more sustained engagement partly driven by reactions to political opponents. For RQ2, they expect left-leaning frames to surge post-Floyd, with observable shifts in narrative engagement. The study aims to clarify how differing attentional profiles and frames affect the potential for social change.

The study situates itself within literature on the attention economy (Simon, 1971; Hendricks and Vestergraad, 2019) and the issue-attention cycle (Downs, 1972), alongside work on social movements’ use of Twitter for rapid mobilization but limited endurance (Tufekci, 2017). Prior research shows polarized attitudes toward BLM (Alfano et al., 2022; Drakulich and Denver, 2022) and that frames strongly affect political perceptions, especially for social movements (Benford and Snow, 2000; Scheufele, 2000; Snow et al., 1986). Studies on online dynamics after mass shootings (Gunn et al., 2018) suggest asymmetric, partisan responsiveness over time. Related BLM Twitter research documents pro- and anti-BLM communities and counter-frames such as #AllLivesMatter and #BlueLivesMatter (Gallagher et al., 2018; Giorgi et al., 2022; Ince et al., 2017; Ray et al., 2017). The paper also draws on work about distributed framing in social media, cross-partisan engagement asymmetries, and algorithmic amplification that may favor right-leaning content, all of which inform expectations about sustained attention and counter-framing.

Data collection: Using a snowball-built list of BLM-related keywords/hashtags (seeded from 2015), the authors collected tweets via the Twitter Streaming API across 2020–2021. They included original tweets and retweets, excluded quote tweets, and amassed ~118.7 million tweets. There was a disruption in API access from July 24 to late August 2020; tweets from this period were included if later retweeted; pure retweets on May 31, 2020 are missing. No bot filtering was performed. Real-world events: Daily counts of U.S. BLM protests were obtained from ACLED, used as exogenous/endogenous variables in time-series models to proxy real-world activity; protest size data were not available. Network and clustering: A retweet network was built where nodes are authors and edge weights count retweets (self-retweets removed). The initial network had ~18M authors and ~98M directed edges (~118M retweets). Edges with weight ≤3 were dropped; the largest connected component (~655k nodes, ~1.7M edges) was retained. Communities were detected with the Leiden algorithm; clusters covering ≥5% of original nodes were kept, yielding major groups labeled Right, Center-Left, and Activist based on top accounts. Time-series models: Daily tweet counts per group and daily protest counts were z-scored. Two models were fit: (1) Naive AR: group tweets predicted from their own lags plus lagged protests (protests exogenous) to estimate responsiveness and decay. (2) Full VAR: vector autoregression where protests are endogenous and groups may influence each other; lag-1 coefficients and impulse response functions (IRFs) were examined over 28 days. Topic modeling: LDA (scikit-learn 1.02) was trained on aggregated preprocessed tweets per author (documents = author-level aggregates) from 2020–2021. Models with k in {3,6,9,…,45} were evaluated via LDA-derived features and linear discriminant analysis classification of clusters; k=24 was selected (clear improvement at 24, diminishing returns thereafter). The fitted model was then used to infer per-tweet topic assignments (by maximum loading); per-group daily topic proportions were analyzed.

- Network clustering revealed three large, polarized communities: Right, Center-Left, and Activist. The Right cluster included prominent conservative media and figures; the Center-Left included mainstream media, civil rights organizations, and politicians; the Activist cluster included self-described activists and left-leaning outlets. - Naive AR results indicated differential attention dynamics: Activists (left) were most responsive to protests but their activity decayed quickly; Right-leaning accounts showed weaker immediate responsiveness to protests but slower decay (more sustained activity); Center-Left was intermediate. • Reported coefficients suggest higher persistence for the Right (e.g., tweet autoregression around r≈0.92 with slower decay ~0.06), faster decay among Activists (e.g., r≈0.76; decay ~0.14), with Center-Left in between (e.g., r≈0.88; decay ~0.10). - Full VAR revealed asymmetries across groups: the Right’s responsiveness to protests was not statistically significant in some specifications, while Activist tweets exhibited a relatively strong positive effect on subsequent protests; interactions suggested the Right often reacts to left-leaning activity rather than to protests directly. - IRFs showed that shocks to tweets produce more sustained effects within the Right cluster than others; shocks to protests generate higher immediate peaks among Activists but with quicker decay. - Topic modeling (24 topics) uncovered distinct pro- and anti-BLM frames and neutral protest/organizational topics. Notably, Topic #2 (e.g., “blm, antifa, terrorist, marxist, riot…”) captured a right-wing “Antifa terrorist” counter-frame. After George Floyd’s murder, right-leaning discourse consolidated around counter-framings that portrayed protests as violent/terroristic, while pro-BLM topics emphasized police brutality, racial injustice, mobilization, and historical identity. - Overall, system shocks (e.g., George Floyd’s murder, protests) elicited rapid, high-amplitude attention from left-leaning users, followed by slower, more persistent right-leaning attention partly driven by reactions to the left.

The findings address RQ1 by demonstrating that different political clusters exhibit distinct attentional dynamics: left-leaning activists rapidly amplify attention to protests but with short-lived effects, whereas right-leaning accounts sustain activity longer, often reacting to left-leaning discourse rather than protests per se. This asymmetry is consistent with prior evidence on cross-partisan engagement and may reflect organized right-leaning media ecosystems that reinforce narratives over time. For RQ2, topic modeling shows that frames diverge across groups and that George Floyd’s murder catalyzed a consolidation of right-wing counter-framing (e.g., branding BLM/Antifa as terrorist), while pro-BLM frames emphasized diagnostic (police brutality, racial injustice), prognostic (protests, action), and motivational (identity, victims) functions. The authors discuss multiple mechanisms: (1) asymmetric cross-partisan influence (Right reacting more to Center-Left); (2) coordinated right-leaning media structures and a “propaganda pipeline” that extend narrative lifespan; (3) distributed framing on social media that enables rapid appropriation and counter-framing; (4) algorithmic amplification and personalization that can create “frame bubbles,” reinforcing certain narratives; and (5) cognitive effects (e.g., “what you see is all there is”) that further entrench frames. These mechanisms help explain why right-leaning narratives can maintain longer attention despite smaller initial surges, highlighting the interplay between network structure, framing, and platform algorithms in shaping public discourse and potential mobilization.

The study shows that proponents and opponents of BLM exhibit systematically different attentional dynamics and framing strategies on Twitter: left-leaning activists generate rapid attention spikes tied to protests, while right-leaning users sustain attention longer, often through counter-framing responsive to left-leaning discourse. The Center-Left plays a notable intermediate role, structurally aligned with activists but showing mixed attentional patterns. Beyond documenting polarization, the work suggests that attentional dynamics themselves are a potential lever for intervention. The authors propose platform-level approaches such as pairing rapidly trending hashtags with alternative or superordinate-identity hashtags to expose users to competing frames early, potentially mitigating frame entrenchment and cross-partisan animus. Future research directions include: leveraging more complete/consistent Twitter data (e.g., academic API) to replicate findings; developing finer-grained subgroup analyses within broad clusters; modeling additional interaction types (replies, quotes) and tracking potential shifts in user affiliations over time; incorporating protest size/severity measures; and examining algorithmic recommendation effects on frame propagation in greater causal detail.

- Data gaps: A ~15-day Streaming API disruption (late July–August 2020) and missing pure retweets on May 31, 2020 may reduce statistical power; resampling the gap was not feasible. - Coarse clustering: Large groupings (Right, Center-Left, Activist) likely contain subgroups; stable, fine-grained subcommunities were difficult to identify reliably. - Interaction scope: Only retweet behavior was used for clustering; quote tweets and replies were excluded from main analyses, limiting insights into endorsement vs. critique and conversational dynamics. - User dynamics: Methods were not designed to detect users changing group affiliation over time; demographic/contextual metadata were not available. - Protest data: Only daily protest counts (not sizes) were used; attendance estimates are scarce and unreliable, limiting inference about protest magnitude. - Bot filtering: No attempt was made to remove bots; conclusions focus on observed dynamics regardless of human/bot mix.