Modular Polyvocality: Community Detection and Semantic Analysis of Civil Society Discourse in English-Language #MahsaAmini Twitter Networks

Abstract

This study investigates how English-language Twitter discourse surrounding the #MahsaAmini protests (October 2022–February 2023) organized into distinct yet interconnected thematic communities within civil society networks. Combining Louvain community detection, sentiment profiling (via VADER), and semantic validation (Jensen–Shannon Divergence, MANOVA, cosine similarity), we map five protest “streams” reflecting modular issue alignments, from Documentation of events involving violence and On-the-Ground Mobilization to Monarchist Advocacy & Diaspora Political Engagement and Anti-IRGC Campaigns. While these communities differ in topical focus, their shared affective tone and overlapping lexicons point to modular polyvocality, a discourse structure characterized by thematic differentiation within broader rhetorical and affective coherence. This pattern represents polyvocal consensus rather than fragmentation. We interpret these findings through the lens of connective action, networked publics, and digital diaspora activism, arguing that the discourse reflects a transnational, modular, and affectively aligned subset of protest-oriented civil society mobilization within constrained information environments. This work contributes to debates on hybrid dissent, algorithmic visibility, and the structure of protest discourse in controlled digital contexts.

Keywords

Mahsa Amini Twitter transnational activism sentiment analysis digital protest network analysis Iran protest discourse

Introduction

Twitter has become a key platform for political communication in contexts with centralized media structures (Kermani, 2023; Michaelsen, 2018). Its accessibility and time-space independence make it a major tool for producing, distributing, and monitoring information, and an arena where public opinion is actively constructed during socio-political transformations. Computational studies of political Twitter discourse demonstrate the value of combining text mining and network analysis to reveal how distinct discursive communities emerge around contested events. Demirel et al. (2024) show that the Hagia Sophia conversion produced differentiated sentiment clusters with distinct emotional registers, structurally similar to the thematic streams in #MahsaAmini discourse while Demirel and Gündüz (2022) demonstrate that COVID-19 Twitter discourse formed identifiable communicative streams with distinct informational functions, illustrating how crises generate predictable structural patterns in networked publics and how social media becomes essential for countering misinformation. In Iran, Twitter enables information sharing, political expression, coalition building, and agenda setting (Kermani, 2023), supporting rapid narrative diffusion across linguistic and national boundaries (Ahmed & Jaidka, 2013). We conceptualize English-language discourse as civil society communication: coordinated expression by non-state actors, including activists, diaspora communities, and advocacy groups operating outside state-controlled media.

The death of Mahsa Amini in September 2022 triggered mobilization that challenged assumptions of digital slacktivism (Kermani, 2023). Expanding from hijab-related debate to broader claims under “Woman, Life, Freedom,” the #MahsaAmini protest gained global visibility through transnational activist and diaspora networks (Khorramrouz et al., 2023). Over 500 million Persian tweets reflect large-scale mobilization (Kermani, 2025), forming what Tufekci (2017) calls a “contested commons” where mourning, memory, and civic action converge. Alongside domestic mobilization (Khatam, 2023), the English-language layer is crucial, produced by activists, diaspora users, and human rights actors, and it reflects strategic reframing for global audiences rather than translation of Persian slogans (Atefinia & Mirnezami, 2025; Michaelsen, 2018). While Farsi tweets focused on domestic issues (Akbari, 2022), English posts targeted international audiences through universal rights discourse and platform visibility logics (Michaelsen, 2018). This globally oriented discourse has a distinct structure not fully captured by existing frameworks. To explain coherence amid thematic diversity, we introduce modular polyvocality.

Conceptualizing Modular Polyvocality

We introduce modular polyvocality to describe a structural pattern in transnational protest discourse: thematic modularity (semi-autonomous discourse communities) coexisting with rhetorical polyvocality (shared affective orientation and overlapping symbolic vocabularies). This extends existing frameworks in three ways. First, while Bennett and Segerberg (2012) “connective action” explains coordination without centralized leadership, modular polyvocality specifies the resulting structure: differentiated thematic streams held together through shared emotional registers. Second, whereas Papacharissi (2015) “affective publics” highlights sentiment as the binding force of diverse actors, our framework emphasizes how affective coherence coexists with strategic thematic specialization rather than requiring uniform messaging. Third, unlike models of digital fragmentation (Barberá et al., 2015) or polarization-centered “echo chambers,” modular polyvocality describes how diversity can enhance collective action by enabling complementary functions, documentation, mobilization, advocacy, and diaspora engagement within a shared moral frame. This pattern is especially visible in transnational activism under political structural constraints, where movements address multiple audiences while navigating censorship and linguistic boundaries (Mercea & Bastos, 2016). Unlike electoral communication that polarizes into binary camps, human rights movements often organize around shared grievances while distributing communicative labor across specialized streams (Bennett & Segerberg, 2012). Modular polyvocality thus captures “strategic complementarity”: differentiated messaging serving distinct tactical roles while reinforcing a common cause (Table 1).

Table 1.

Operational Definitions for Discourse Architecture

Term	Definition
Community	A discrete cluster of users identified by the Louvain algorithm based on retweet interactions.
Stream	A higher-order thematic grouping of multiple communities sharing a common rhetorical focus.
Module	The functional unit of discourse within a stream that addresses specific grievances or goals.
Modular polyvocality	A discourse structure characterized by thematic differentiation (modules) coexisting within a broader rhetorical and affective consensus.

Literature Review

Transnational Activism and the Strategic Use of English

Transnational activism literature shows how local struggles are embedded in global networks (Piper & Uhlin, 2004), often through the “boomerang effect,” where external pressure compensates for constrained domestic channels (Keck & Sikkink, 1998). English functions as a strategic bridge, enabled by digital platforms and “serial transnational activists” who translate local testimony into globally legible frames, a practice extending beyond diaspora actors to domestic activists seeking international visibility (Mercea & Bastos, 2016).

The Iranian Twittersphere as a Contested Transnational Arena

Since the 2009 Green protests, the Iranian Twittersphere has been a contested space for political communication (Kermani, 2023). The #MahsaAmini protest reframed it as strategic action beyond “slacktivism” critiques (Tufekci, 2017). Activism operates under information control and filtering (Sohrabi-Haghighat, 2011) as well as “informational saturation” driven by bots and noise (Kermani, 2025), within which emotional expression shifts into citizen documentation (Rosenbaum & Bouvier, 2020), combining expressive and strategic communication (Kermani, 2023) and reinforcing digital resilience and transnational solidarity (Khorramrouz et al., 2023).

Theoretical Framework: Networked Publics, Gatekeeping, and Framing

This study integrates three theoretical domains. First, connective action explains how loosely coordinated discourse “streams” emerge through personalized digital networks rather than hierarchical organization (Bennett & Segerberg, 2012). Second, these streams function as networked publics where narratives develop outside state-affiliated media (Boyd, 2010; Papacharissi and de Fatima Oliveira, 2012); traditional gatekeeping is replaced by “networked gatekeeping” as hashtags, retweets, and platform algorithms shape visibility (Meraz & Papacharissi, 2013; Papacharissi, 2015), capable of both unifying and fragmenting movements (Meraz, 2017) while remaining vulnerable to bots and misinformation (Golkar, 2015; Kermani, 2025). Third, we draw on digital diaspora activism (Ghorashi & Boersma, 2009; Sohrabi-Haghighat, 2011): English-language tweets function as strategic public diplomacy rather than direct translation, as diasporic and allied actors reframe narratives for global audiences, producing a “contested commons of visibility” (Tufekci, 2017).

From Fragmentation to Modular Polyvocality: Theorizing Discourse Structure

Across scholarship on digitally mediated protest, collective coherence under semantic diversity is explained through a limited set of mechanisms. A first group, networked framing, affective publics, connective action, and hybrid movement solidarity, locates coherence in emotional alignment, where shared affect binds participants despite divergent narratives (Meraz & Papacharissi, 2013; Papacharissi, 2016; Pavan & Rapini, 2022; Rosenbaum & Bouvier, 2020; Stewart & Schultze, 2019). A second group explains coherence through organizational coordination under controlled socio-political atmosphere, treating semantic diversity as destabilizing or as a product of controlling strategies (Kermani, 2025; Khatam, 2023). A third strand, aspect-based sentiment analysis, emotion lexicons, and structural approaches such as random matrix theory, maps emotional or structural variation without theorizing how coherence persists under fragmentation (Haralabopoulos & Simperl, 2017; Mohammadi et al., 2022).

Taken together, these literature studies explain persistent coherence either through affective alignment, organizational coordination, or not at all. Affect’s role as an autonomous mechanism synchronizing heterogeneous communities remains under-specified (Papacharissi, 2016; Rosenbaum & Bouvier, 2020). This limitation surfaces in recurring empirical tensions: online–offline disjunctions (Pavan & Rapini, 2022), doubts about the effectiveness of connective movements (Michelen & Glasius, 2018; Rosenbaum & Bouvier, 2020), participant anonymity and fragmentation (Etemadifard & Khazaei, 2021; Rosenbaum & Bouvier, 2020; Stewart & Schultze, 2019), and difficulties sustaining coherence in polarized networks (Kermani, 2025). What remains under-theorized is not polyvocality or affective engagement, but the relational condition in which they coexist without collapsing into fragmentation, how affective alignment operates across modular structures (Papacharissi, 2016), how semantic divergence is contained, and how coherence persists without shared frames or centralized coordination.

Our Contribution

Existing theories either attribute unity to affective alignment, locate it in organizational capacity, or merely describe fragmentation, none theorizes the condition under which heterogeneous discourse remains collectively synchronized across modular communicative structures. This study addresses that explanatory gap by conceptualizing modular polyvocality as the condition in which multiple, internally coherent discursive modules coexist, remain affectively attuned, and contribute to collective persistence without requiring discursive convergence or hierarchical integration.

Our analysis integrates computational network methods with qualitative interpretation to answer three interrelated research questions.

RQ1

Do unsupervised methods reveal thematic clustering in this discourse? Method: Louvain community detection (Section 2.2) applied to monthly retweet networks, measuring clustering strength through modularity scores. Output: Network visualizations, modularity statistics, and community membership assignments.

RQ2

Do identified clusters show thematic differentiation? Methods: (a) Qualitative coding (Section 2.5); (b) lexical profiling via G² (Section 2.6.1); (c) semantic validation through word embeddings (Section 2.7). Output: Five-stream typology validated through intercoder reliability, keyword profiles, and semantic coherence tests.

RQ3

How do thematic patterns evolve over time? Methods: Temporal segmentation (Section 2.7), semantic drift analysis (Section 2.8.2), and period-based comparison tests. Output: Month-to-month vocabulary shifts, hierarchical clustering revealing meta-streams, and chi-square tests of period differences.

Cross-Cutting Analyses

Bot detection (section 2.3), sentiment analysis (Section 2.9), and topic divergence (Section 2.6.2) (Figure 1).

Figure 1.

Visualization of the analytical pipeline, showing how raw tweets flow through successive stages, creating a multi-layered validation structure where findings must be consistent across computational and interpretive methods

Materials and Methods

Data Collection and Preprocessing

English-language tweets containing #MahsaAmini and related hashtags were collected via the Twitter API over 6 months (October 2022–February 2023), covering the movement’s peak transnational visibility (Kermani, 2025). The focus on English tweets reflects their role as a channel for global advocacy rather than user geography.

Preprocessing followed a multi-stage pipeline. Regex removed hyperlinks, mentions, and non-textual symbols while preserving core contextual information (Haralabopoulos & Simperl, 2017). Non-ASCII characters and emojis were encoded and filtered to retain English-language semantic structure. Stopwords were removed using NLTK (Bird et al., 2009) to optimize the corpus for clustering, embeddings, and keyword extraction. Mixed-language tweets containing both English and Persian script were retained but treated as English-language documents; Persian-script tokens were removed alongside other non-ASCII characters during filtering, meaning only the English-language portion contributed to downstream lexical and semantic analysis. This approach preserves the strategic English-language layer of the discourse while acknowledging that meaning carried exclusively in Persian script is not captured, a limitation discussed in Section 4.7.

A Temporal Text-Matching algorithm was applied to detect and group duplicate content, ensuring accurate inference of retweet relationships and network robustness even when API metadata was incomplete (Figure 2).

Figure 2.

Top ten most prevalent hashtags during the studied period

Community Detection Algorithm

Monthly retweet networks were constructed with users as nodes and retweet relations as directed edges, capturing information flow and alignment (Castillo, 2016). To incorporate both behavioral alignment (retweet relations) and topical alignment (hashtag co-occurrence), retweet and co-hashtag edges were merged into a unified graph prior to community detection, producing more robust community structures reflecting multiple dimensions of discursive alignment (Castillo, 2016; Mohammadi et al., 2022). All edges were treated as unweighted binary ties; while weighted networks capture interaction intensity, unweighted representations are more appropriate in protest settings where connectivity itself indicates alignment (Farzam et al., 2023). Mention networks were excluded due to susceptibility to manipulation in politically contested contexts (Kermani, 2025; Tufekci, 2017). Where API metadata was incomplete due to deletions or suspensions, the Temporal Text-Matching algorithm reconstructed missing retweet relations by grouping duplicate content (Olteanu et al., 2019).

Community detection used the Louvain modularity optimization algorithm (Farzam et al., 2023; Mohammadi et al., 2022), selected for scalability in large networks. Each monthly network was run with 10 random initializations, retaining the highest-modularity partition to reduce sensitivity to local optima, with a resolution parameter of 1.0 to balance granularity and over-fragmentation (Mohammadi et al., 2022). The resulting clusters formed the basis for downstream semantic and lexical analyses (Campbell et al., 2013).

Bot Flagging

To preserve analytical validity in an environment with coordinated inauthentic activity, we constructed a composite bot_score integrating five validated dimensions: temporal activity (Al-azawi & Safaa, 2022), account metadata (Varol et al., 2017), network behavior (Castillo et al., 2011), content style (Kosmajac & Keselj, 2019), and coordinated surges indicating botnet activity (Davis et al., 2016). Features were weighted by prior reliability (Varol et al., 2017), producing a normalized score (0–1). A conservative threshold of 0.7 was applied to reduce false positives in politically sensitive contexts (Keller & Klinger, 2019), where misclassification of activists could distort findings (Pacheco et al., 2021; Rauchfleisch & Kaiser, 2020).

Bot detection followed a two-stage design. In stage one, tweets from flagged accounts were removed from all content-level analyses (G² profiling, Jensen–Shannon divergence, semantic validation, and VADER sentiment analysis). In stage two, flagged accounts were retained in the network graph to preserve structural integrity, as removal would distort community topology (Pacheco et al., 2021; Rauchfleisch & Kaiser, 2020), a trade-off discussed in Section 4.7. This separation follows established computational social science practice (Castillo et al., 2011; Varol et al., 2017).

Of 40,162 tweets, 11,276 (28.08%) were classified as bot-generated and excluded from content analysis. Bot activity varied across months (1.5%–14.5% in October 2022; 1.5%–9% in February 2023), with a sharp spike in December 2022 where one community reached ∼45%, coinciding with the Final judicial sentences carried out by the state against Karami and Hosseini, consistent with Kermani (2025) concept of reactive “informational saturation.” Across the five thematic streams, mean bot scores were uniformly low (0.033–0.195). Stream five exhibited the highest average bot score (0.195), while Stream four showed the lowest (0.033). A one-way ANOVA revealed no statistically significant differences in bot activity across streams (F (4, 21) = 0.68, p = .626), indicating that automated activity was distributed broadly rather than concentrated in any specific discourse stream. This supports the interpretation that the observed modular structure reflects organic human discourse rather than an artifact of coordinated automated amplification.

Stream Mapping and Reliability Assessment

Community detection alone does not reveal thematic meaning (Braun & Clarke, 2006). We applied an interpretive stream mapping framework linking structural communities to thematic categories through qualitative coding (Farzam et al., 2023; Kermani, 2025), treating streams as analytical constructs rather than fixed discourse boundaries (O'Connor et al., 2010).

The largest communities per monthly network were selected based on size and modularity score (Mohammadi et al., 2022). Saturation testing at five sampling levels (10, 25, 50, 75, and 100 tweets per community) showed thematic stabilization between 50 and 75 tweets (6.1 dominant themes at both levels; 6.0 at 100), consistent with established saturation logic (Guest et al., 2006). Our 75-tweet threshold, applied across 34 communities (2,550 tweets), achieved Krippendorff’s α = 0.74, supporting stable and replicable community-level representations. Adequacy of this strategy is further confirmed by convergence between 75- and 100-tweet samples alongside validation from G² profiling, Jensen–Shannon divergence, MANOVA, and sentiment analysis.

Manual coding followed three steps. First, key users were identified by retweet volume for network description only, not sampling (Castillo et al., 2011). Second, 75 tweets per community were randomly sampled to avoid visibility bias from algorithmic amplification (Kermani, 2025; Olteanu et al., 2019), using a fixed seed (random_state = 42) for reproducibility. Third, sampled tweets were coded using an iterative thematic framework and codebook (Appendix A), following established computational qualitative methods (Braun & Clarke, 2006; Freelon & Wells, 2020).

Reliability Assessment

Four independent coders with Iranian cultural background and expertise in political geography and political sciences, two Master’s graduates, and two PhD candidates assessed a stratified sample of 10 communities (2 per month) to ensure temporal representativeness (Castillo, 2016; Pavan & Rapini, 2022). Working independently with the same codebook (Freelon & Wells, 2020), they achieved Krippendorff’s α = 0.74, exceeding conventional thresholds for acceptable agreement (Krippendorff, 2013). Iranian coders’ cultural competency proved valuable for interpreting embedded Persian terms and implicit political meanings (Michaelsen, 2018), while disciplinary diversity reduced analytical bias (Braun & Clarke, 2006). This hybrid method, automated clustering plus interpretive labeling with validated reliability, enhances robustness while acknowledging that findings represent one possible interpretation of underlying clustering patterns (Gerbaudo, 2012; Tufekci, 2017), with the five categories functioning as interpretive frameworks rather than objective discourse boundaries (Papacharissi & de Fatima Oliveira, 2012) (Table 2).

Table 2.

Saturation analysis showing average number of dominant themes detected at five sample sizes across monthly retweet communities (October 2022–February 2023)

Lexical Characterization and Topic Divergence

Lexical Characterization (G² Keyword Profiling)

To identify stream-specific vocabulary, we applied the log-likelihood ratio test (G²), a standard corpus linguistics method for detecting overrepresented terms (Dunning, 1993). For each stream, a 2 × 2 contingency table compared word frequencies within the stream against all others; terms with G² ≥ 6.63 (p < 0.01) and minimum frequency thresholds were retained. The top 30 terms per stream served as lexical fingerprints capturing distinct rhetorical and thematic emphases (Ahmed & Jaidka, 2013; Haralabopoulos & Simperl, 2017).

Topic Divergence (Jensen–Shannon Divergence)

To assess thematic separation, we combined probabilistic topic modeling with information-theoretic divergence. LDA was applied via Gensim’s LdaModel (Řehůřek & Sojka, 2010) to identify latent thematic structure, representing documents as topic mixtures and topics as word distributions. The model was trained jointly on all five streams to ensure a shared topic space, with k = 10 topics, 10 passes for convergence, a fixed seed (random_state = 42), and vocabulary filtered by frequency thresholds (no_below = 10, no_above = 0.5). Document-level topic distributions were averaged per stream into a single representative vector (“topical DNA”) using Gensim’s sparse2full function. Pairwise Jensen–Shannon Divergence (JSD) was then computed between stream-level vectors. JSD is a symmetric, bounded measure (0–1) of distributional divergence (Tan & D’Souza, 2025) that avoids the asymmetry and zero-probability issues of KL divergence, making it suitable for partially overlapping vocabularies. Together, LDA-based JSD and G² profiling provide complementary measures capturing both deep topic distributions and surface lexical differentiation.

Semantic Validation

Word Embedding Approach

We used pre-trained fastText English embeddings (cc.en.300.vec), representing words as high-dimensional vectors based on contextual similarity (Khorramrouz et al., 2023). The top 30 G²-identified keywords from each stream were converted into vectors using gensim.KeyedVectors.

Semantic Coherence Tests

Three procedures validated stream vocabulary coherence: (a) Centroid Construction, embedding vectors of each stream’s top keywords were averaged to create a representative “semantic center of gravity”; (b) Seed Alignment, stream keywords were compared to predefined thematic seed centroids using cosine similarity; (c) Label Alignment, each stream centroid’s proximity to a manually constructed reference label was measured. This confirmed alignment between stream content and label, especially in narrower streams like Incident Documentation (cosine = 0.75), while broader streams showed lower scores due to lexical diversity.

Semantic Space Mapping via Multidimensional Scaling (MDS)

MDS was applied to cosine similarity matrices derived from top keyword embeddings, projecting high-dimensional word-vector relationships into 2D space to enable visual inspection of community clustering. A two-sample t-test on Euclidean distances in MDS-reduced space evaluated intra-vs. inter-stream coherence.

Statistical Validation (MANOVA Tests)

Two MANOVA tests using MDS coordinates were conducted: the full embedding space yielded no significant group separation (p = 0.744), likely due to overlapping vocabulary; the topword embedding space showed significant differences (Wilks’ λ = 0.1789, p < 0.001), affirming stream separation when lexical scaffolding is used. All labeling was completed prior to validation to avoid circularity.

Temporal and Hierarchical Analyses

Hierarchical Clustering: Meta-Streams

Agglomerative hierarchical clustering on top-word frequency vectors identified higher-order “meta-streams,” revealing thematic alignments (e.g., early mobilization vs. geopolitical advocacy) beyond initial community detection (Braun & Clarke, 2006).

Semantic Drift: Month-to-Month Lexical Change

Semantic drift was measured by calculating Euclidean distances between monthly lexical centroids, classifying streams as persistent (stable vocabulary), emergent (appearing later), or sporadic (rapidly changing) (Castillo, 2016), capturing narrative adaptation to protest dynamics or algorithmic shifts (Pavan & Rapini, 2022). A chi-squared test on normalized token frequencies between two periods (Oct–Dec 2022 vs. Jan–Feb 2023) confirmed a significant shift from grassroots terms to international advocacy frames, coinciding with the Final judicial sentences carried out by the state against Karami and Hosseini. A co-occurrence matrix aligning top terms with protest events (e.g., “Zahedan Black Friday”) further revealed how streams semantically reoriented around symbolic flashpoints (Castillo, 2016).

Sentiment Analysis Using VADER

Affective variation was assessed using VADER (Valence Aware Dictionary and sentiment Reasoner), a lexicon- and rule-based model validated for social media text (Hutto & Gilbert, 2014). For each stream, VADER derived averaged positive, negative, neutral, and compound scores; the compound score served as the primary metric. While all streams exhibited negative affective valence, intensity varied, from most negative (Stream 3, Anti-IRGC Campaigns: −0.257) to most neutral (Stream 4, Monarchist Advocacy & Diaspora Political Engagement: −0.028). One-way ANOVA confirmed significant variation across mean compound scores, supporting the hypothesis of emotionally distinct discourse formations reflecting underlying strategic differences.

Results

Network Modularity and Community Structure

Monthly retweet networks (October 2022–February 2023) were analyzed using the Louvain algorithm (resolution = 1.0). Despite network size declining from ∼9,000 users in October to ∼3,100 in February, modularity remained stable and high (0.6137–0.7261; avg. 0.6586), with 17–26 communities detected per month, indicating a strongly clustered, non-random network structure.

Thematic Stream Reliability Validation

Structural communities were interpreted through qualitative stream labeling, yielding five discourse formations: (1) Documentation of events involving violence, (2) On-the-Ground Mobilization, (3) Anti-IRGC Campaigns, (4) Monarchist Advocacy & Diaspora Political Engagement, and (5) Broader Opposition Narratives & International Lobbying, inductively defined via close reading of 75 tweets per community. Inter-coder reliability across 10 stratified communities reached Krippendorff’s α = 0.74, supporting the stability of the typology (Freelon & Wells, 2020; Rauchfleisch & Kaiser, 2020) (Figure 3; Tables 3 and 4).

Figure 3.

3D visualization of the retweet and co-hashtag network in January, illustrating modular community structure and inter-stream clustering

Table 3.

Summary Statistics of Monthly Unified Retweet and Co-Hashtag Networks

Month	Nodes (users)	Edges (retweets, co-hashtags)	Modularity	Communities detected	Top 5 community sizes (users)
Oct 2022	8,939	223,646	0.6244	26	900, 644, 540, 378, 152
Nov 2022	8,129	203,948	0.7261	25	549, 293, 280, 218, 168
Dec 2022	7,271	187,990	0.6137	23	737, 398, 372, 224, 164
Jan 2023	3,862	92,584	0.6475	19	444, 423, 356, 323, 36
Feb 2023	3,152	86,633	0.6815	17	354, 335, 214, 125, 97

Table 4.

Inter-Rater Agreement Across Four Coders in Thematic Stream Labeling, Using Krippendorff’s Alpha

Stream	Stream label	Perfect agreement (4/4) (%)	Majority agreement (3/4) (%)	Tie (%)	No agreement (all different) (%)	Overall agreement rate (%)	Krippendorff’s alpha
1	Documentation of events involving violence	70	30	0	0	100	0.85
2	On-the-Ground Mobilization	50	40	10	0	90	0.78
3	Anti-IRGC Campaigns	60	30	10	0	90	0.82
4	Monarchist Advocacy & Diaspora Political Engagement	30	50	10	10	80	0.65
5	Broader Opposition Narratives & Lobbying	40	50	10	0	90	0.72
–	Overall (all streams)	50	40	7	3	90	0.74

The Contested Information Environment: Bot-Like Activity Across Streams

Bot-like activity was measured using a composite 0–1 bot score. Mean scores across streams were low (0.033–0.195), with ANOVA showing no significant differences between them (F = 0.682, p = 0.626), indicating evenly distributed rather than stream-targeted automated activity. The modular structure therefore reflects organic human discourse rather than coordinated automated interference (Tables 5 and 6).

Table 5.

Mean Bot-Likelihood Scores by Discourse Stream, Based on Composite Behavioral Indicators

Stream	Average bot score
Stream 5: Broader opposition narratives	0.195
Stream 1: Documentation of events involving violence	0.162
Stream 3: Anti-IRGC Campaigns	0.086
Stream 2: On-the-Ground Mobilization	0.070
Stream 4: Monarchist Advocacy & Diaspora Political Engagement	0.033

Table 6.

Bot Activity Summary Across Monthly Communities and Thematic Streams

Period	Bot proportion range (across communities)	Notable observations
October 2022	6.0%–14.5%	Baseline activity; movement peak
November 2022	1.5%–27.5%	Moderate variation
December 2022	7.0%–45.0%	Spike in one community coinciding with executions
January 2023	6.0%–31.5%	Elevated; The period of Karami–Hosseini judicial sentences enforcement
February 2023	1.5%–9.0%	Sharp decline as movement intensity reduced
Overall	28.08% of 40,162 tweets flagged	Evenly distributed across streams (ANOVA p = .626)

Mean bot scores per stream ranged from 0.061 (Stream 4) to 0.225 (Stream 1). Given the limited number of community observations per stream (n = 1–4), a formal inferential test is underpowered; the descriptive pattern nonetheless indicates no concentration of bot activity in any single discourse stream, consistent with the manuscript’s argument for organic modular structure.

Lexical Distinctions Across Streams

The top 30 distinctive keywords per stream were extracted via G² profiling (G² ≥ 6.63, p < 0.01), revealing distinct lexical profiles: Stream 1 (Incident Documentation): killed, bloody, funeral, martyr, violence; Stream 2 (On-the-Ground Mobilization): murder, twitter, chant, capital, liberation; Stream 3 (Anti-IRGC Campaigns): sanctions, IRGC, regime, accountability; Stream 4 (Monarchist Advocacy & Diaspora Political Engagement): Reza, Pahlavi, monarchy, restoration, dynasty; Stream 5 (Broader Opposition Narratives & Lobbying): parliament, resolution, embassy, diplomatic. While terms like “freedom” and “justice” appeared across streams, their statistical overrepresentation and contextual associations varied, confirming each stream’s distinct lexical profile (Table 7).

Table 7.

Top Discriminative Keywords Per Stream Ranked by Normalized G² Values, Indicating Lexical Distinctiveness

Stream	Word	g_score	p_value	is_significant	observed_in_stream	expected_in_stream	observed_in_total
Stream 1	Armitaabbasi	3550.032	0	TRUE	3201	1674.047	3320
Stream 1	Foxnews	1636.124	0	TRUE	1198	604.0686	1198
Stream 1	Workers	1512.33	0	TRUE	1108	558.6878	1108
Stream 1	Saytheirnames	1431.206	0	TRUE	1049	528.9382	1049
Stream 1	Cnn	1354.807	########	TRUE	3610	2385.012	4730
Stream 2	Murder	3468.198	0	TRUE	1105	230.4759	1105
Stream 2	Twitter	2096.043	0	TRUE	670	139.7456	670
Stream 2	Dec	1592.544	0	TRUE	510	106.3735	510
Stream 2	Chant	1561.098	0	TRUE	500	104.2877	500
Stream 2	Capital	1438.484	0	TRUE	461	96.15329	461
Stream 3	irgc	8423.032	0	TRUE	4011	819.1354	7934
Stream 3	Terrorist	2766.753	0	TRUE	1368	281.4423	2726
Stream 3: s	Josepborrellf	2068.24	0	TRUE	615	79.18791	767
Stream 3	Organization	1897.567	0	TRUE	419	43.25911	419
Stream 3: s	Tortured	1153.67	########	TRUE	399	58.8489	570
Stream 4	irgc	4527.631	0	TRUE	1941	304.7643	7934
Stream 4	Walkoutiri	1503.496	0	TRUE	406	32.95787	858
Stream 4	Baraye	930.3908	########	TRUE	144	5.531391	144
Stream 4	february	859.8293	########	TRUE	214	15.28815	398
Stream 4	Song	852.0005	########	TRUE	132	5.070442	132
Stream 5	november	3344.057	0	TRUE	1624	341.6535	2875
Stream 5	Uprising	2370.573	0	TRUE	1068	210.2209	1769
Stream 5	Iranregimechange	2090.792	0	TRUE	1048	225.1943	1895
Stream 5	Locals	1236.473	########	TRUE	292	34.70012	292
Stream 5	Rally	1048.659	########	TRUE	248	29.47133	248

Topic Divergence (JSD Matrix)

Pairwise Jensen–Shannon Divergence scores computed from LDA-derived topic distributions were low (0.0626–0.1550), indicating significant topic overlap and a shared vocabulary of protest and resistance across all thematic groupings (Tables 8 and 9).

Table 8.

Jensen–Shannon Divergence Matrix Showing Semantic Distances Between Identified Streams

	Stream 1	Stream 2	Stream 3	Stream 4	Stream 5
Stream 1	0	0.062589	0.078523	0.134523	0.154983
Stream 2	0.062589	0	0.063246	0.124794	0.115871
Stream 3	0.078523	0.063246	0	0.143557	0.125602
Stream 4	0.134523	0.124794	0.143557	0	0.098523
Stream 5	0.154983	0.115871	0.125602	0.098523	0

Table 9.

Cosine Similarity Between Manual Stream Assignments and Computationally Identified Dominant Themes

Stream	Label	Seed keywords	Cosine similarity
1	Documentation of events involving violence	Prison, blood, justice, riot, force, media, victim, jail, death, shoot, support	0.7571
2	On-the-Ground Mobilization	People, call, protest, revolution, march, street, police	0.6144
3	Anti-IRGC Campaigns	Sanction, irgc, international, end, close, embassy, regime, ban, campaign , government	0.5881
4	Monarchist Advocacy & Diaspora Political Engagement	United, together, national, country, Iran, change, support, pahlavi, prince	0.5347
5	Broader Opposition Narratives	Democracy, party, groups, opposition, unity, activist, media, voice	0.4876

Stream Validation—Theme–Centroid Alignments

Pre-trained fastText embeddings assessed semantic coherence of each stream’s top 30 keywords. MDS plots revealed communities clustering spatially by designated stream, with particularly tight groupings for Streams 1 and 3. A two-sample t-test on Euclidean distances confirmed that communities within the same stream were significantly closer than those from different streams (t (274) = −3.48, p = 0.0008), supporting thematic differentiation despite lexical overlap.

Intra-stream distances were significantly smaller than inter-stream distances (M = 0.3430 vs. M = 0.4582), confirming tighter semantic coherence within streams than across them (Figures 4 and 5).

Figure 4.

Two-dimensional MDS plot of community-level linguistic similarity based on cosine distances between top-term vectors

Figure 5.

Distribution of MDS distances within versus between labeled streams, suggesting intra-stream coherence

Meta-Streams and Temporal Drift

Agglomerative hierarchical clustering of thematic centroids produced higher-order “meta-streams”: a two-cluster solution grouped Streams 1 and 2 against Streams 3, 4, and 5. Semantic drift analysis classified streams as Persistent (Streams 1 and 4 stable vocabularies) or Variable (Streams 2, 3, and 5 higher monthly lexical change), measured by Euclidean distances between monthly lexical centroids based on averaged fastText embedding vectors (Figures 6 and 7).

Figure 6.

Hierarchical clustering dendrogram of retweet communities based on cosine similarity of top tokens, illustrating higher-order groupings

Figure 7.

Temporal semantic drift for each stream measured by Euclidean distance across months, indicating thematic evolution

MANOVA Findings: Statistical Differentiation of Streams

MANOVA on MDS coordinates was conducted following assumption checks: homogeneity of covariance was supported via Levene’s test (W = 0.72–2.09, p = .12–.59), with Shapiro–Wilk tests indicating mostly normal distributions with minor violations in very small groups (n = 3–4). Given these constraints, Pillai’s Trace (V = 0.864, p < .001) is reported alongside Wilks’ λ for robustness (Tabachnick & Fidell, 2013).

The two MANOVA tests produced contrasting outcomes, no significant separation in the full embedding space, but significant differences in the topword embedding space interpreted not as contradiction but as methodological insight. The null result likely reflects genuine thematic overlap and shared protest lexicon (Papacharissi & de Fatima Oliveira, 2012; Tufekci, 2017), while the significant result suggests curated lexical features enhance detectability of structural differences, though potentially amplifying separation through analytical design choices (O'Connor et al., 2010; Milan, 2015; Olteanu et al., 2019). Overall, results support the analytical utility of the thematic framework while demonstrating that statistical separability is conditioned by methodological decisions in complex discourse environments (Bennett & Segerberg, 2012; Braun & Clarke, 2006; Jackson et al., 2020; Meraz & Papacharissi, 2013) (Table 10).

Table 10.

Comparison of MANOVA Results for Top-Word and Embedding-Based Representations of Stream-Level Discourse

Test	Streams included	Wilks’ lambda	Pillai’s Trace	Hotelling-Lawley	Roy’s root	F-value	p-Value	Statistical significance	Interpretation
Topwords-based MANOVA	24	0.1789	0.8644	4.3458	4.2893	6.1378	0.0001	Yes (p < 0.05)	Streams differ significantly in MDS space of top keywords
Embedding-based MANOVA	24	0.7683	0.2338	0.2988	0.2892	0.6338	0.744	No (p > 0.05)	No significant differences in full embedding MDS space

Stream Evolution Over Time

Monthly keyword analysis reveals lexical evolution aligned with key protest events. Stream 1 shifted from “killed” and “massacre” during the Zahedan Black Friday events (October) to “martyr” and “funeral” during 40-day memorials (November), while Stream 3 peaked in “execution” and “hanged” following the Final judicial sentences carried out by the state against Karami and Hosseini (January). A chi-square test confirms a significant thematic shift between Period 1 (Oct–Dec 2022) and Period 2 (Jan–Feb 2023) (p < 0.001): Period 1 was dominated by on-the-ground reporting (8.7%) and mourning terms, whereas Period 2 shows a rise in sanctions and lobbying terms from 5.6% to 11.8%, marking a transition from localized resistance discourse to transnational advocacy and forming two temporally distinct meta-streams (Figure 8; Tables 11, 12 and 13).

Figure 8.

Top 15 discriminative tokens across Period 1 and Period 2, normalized by token frequency, indicating evolving protest narratives

Table 11.

Most Frequent Stream-Specific Terms During Three Critical Protest Moments, Highlighting Event-Driven Discourse Peaks

Stream	Event	Month year	Top co-occurring terms
Stream 1	Evin prison fire	2022-10	Prisoners, tehran, heard
Stream 1	Final judicial sentences carried out by the state against Karami and Hosseini	2023-01	Arrested, corruption, mehdimohammadifard
Stream 1	Mohsen Shekari execution	2022-12	Saytheirnames, protester, arrested
Stream 1	Sharif protest	2022-10	Silent, remaining, befairun
Stream 1	Zahedan black friday	2022-10	killed,زاهدان (Zahedan), bloody
Stream 2	Evin prison fire	2022-10	Tehran, prisoners, heard
Stream 2	Mohsen Shekari execution	2022-12	Prison, arrested, zahedan
Stream 2	Sharif protest	2022-10	Tehran, chanting, school
Stream 2	Zahedan black friday	2022-10	Killed, sanandaj, زاهدان (Zahedan),
Stream 3	Enforcement of Final judicial sentences of Karami and Hosseini	2023-01	Arrested, prison, corruption
Stream 5	Evin prison fire	2022-10	Prisoners, tehran, citizens
Stream 5	Sharif protest	2022-10	Tehran, oct, death
Stream 5	Zahedan black friday	2022-10	Crime, humanity, human

Table 12.

Shift in Dominant Vocabulary Between Period 1 (Oct–Dec) and Period 2 (Jan–Feb), Showing Thematic Realignment

Period 1 (Oct-Dec 2022)	Freq (P1)	Period 2 (Jan-Feb 2023)	Freq (P2)
Protest	2456	Sanctions	1998
Killed	1892	International	1756
Street	1634	Execution	1634
Police	1445	Pressure	1523
Funeral	1289	Government	1445

Table 13.

Chi-Square Test to Assess the Difference on Top Words Counts Between Two Periods

Theme	count_period_1	count_period_2	norm_freq_diff	prop_period_1	prop_period_2	chi2	Significant
International sanction	58,678	24,690	0.061929	0.056222	0.118151	0	TRUE
Detention/custody	52,695	14,671	0.019717	0.05049	0.070206	########	TRUE
Time	8261	1911	0.00123	0.007915	0.009145	1.17E-08	TRUE
Political	6324	1088	−0.00085	0.006059	0.005206	3.76E-06	TRUE
Media	6175	926	−0.00149	0.005917	0.004431	1.74E-16	TRUE
Religion	4233	503	−0.00165	0.004056	0.002407	4.53E-29	TRUE
Mourning	22,628	3577	−0.00456	0.021681	0.017117	2.39E-40	TRUE
Other	794,130	151,337	−0.03669	0.760896	0.724204	########	TRUE
On the ground	90,554	10,267	−0.03763	0.086764	0.049131	0	TRUE

Affective Signatures: Sentiment Analysis of Discourse Streams

VADER sentiment analysis revealed distinct affective signatures across streams. While all streams exhibited negative average compound scores, a one-way ANOVA confirmed a statistically significant difference in mean compound sentiment scores across the five discourse formations, F (4, 116,914) = 22.58, p < .001, η² = 0.0008. To determine which specific streams differed, a Tukey’s HSD post-hoc test was conducted. The post-hoc analysis indicated that all pairwise comparisons between the streams were statistically significant (p < .001). Specifically, Stream 3 (Anti-IRGC Campaigns) was significantly more negative (Avg Compound = −0.257, SD = 0.35) than all other streams, functioning as the most denunciatory cluster. Conversely, Stream 4 (Monarchist Advocacy) was significantly more neutral (Avg Compound = −0.028, SD = 0.35) compared to the baseline of the other streams (p < .001). From a practical standpoint, the remarkably small effect size (η² = 0.0008) despite high statistical significance empirically supports our claim of “affective alignment”: the streams differ statistically in their emotional registers, but remain functionally consistent in their overall negative tone across the modular discourse architecture (Tables 14 and 15).

Table 14.

Average VADER Sentiment Scores (Positive, Negative, Neutral, Compound) Across Thematic Streams

Stream	Avg compound	Std dev (compound)	Avg negative	Avg neutral	Avg positive
Stream 3	−0.257	0.34	0.148	0.758	0.094
Stream 1	−0.178	0.35	0.125	0.787	0.088
Stream 2	−0.126	0.35	0.117	0.788	0.095
Stream 5	−0.076	0.35	0.108	0.793	0.099
Stream 4	−0.028	0.35	0.104	0.811	0.085

Table 15.

ANOVA Results for VADER Sentiment Scores

Metric	F-statistic	Interpretation
Compound	22.58	Very large: Strong evidence that streams differ in overall sentiment polarity
Negative	11.20	Also large: Streams differ in how negative their tweets are, significantly
Neutral	9.23	Significant: Streams vary in use of neutral language (e.g., monarchist higher neutrality)
Positive	3.16	Modest: weak evidence, could be noise; differences in positivity across streams are not very strong

Discussion: Modular Polyvocality in Transnational Digital Activism

This study’s central contribution is the empirical identification of modular polyvocality as a distinct organizational pattern in transnational protest discourse. The English-language #MahsaAmini discourse exhibits high structural modularity (0.61–0.73), clear lexical differentiation (G² stream-specific vocabularies), and measurable thematic boundaries (MANOVA separation in topword space, t (274) = −3.48, p < 0.001), while maintaining affective coherence and semantic proximity (Jensen–Shannon Divergence: 0.06–0.15), defining features of modular polyvocality: distinct thematic clusters coexisting with rhetorical and emotional cohesion. This structure is neither fragmentation nor homogeneity, but strategic complementarity, where streams perform specialized functions, documentation, mobilization, sanctions advocacy, diaspora engagement, and Broader Opposition Narratives & Lobbying while reinforcing a shared cause through distributed rhetorical labor. Its persistence over 5 months suggests strategic adaptation rather than algorithmic artifact.

Using a hybrid computational–interpretive approach, we identify five stable streams with lexically clear but semantically porous boundaries. This challenges the fragmentation–unity binary in early social media theory (Barberá et al., 2015; Sunstein, 2018) and aligns with differentiated information flows (Jackson et al., 2020), specifying modular polyvocality as the enabling mechanism. The structure reflects connective action logic, mobilization through circulating symbolic content rather than centralized coordination (Bennett & Segerberg, 2012)—and exemplifies networked publics, where platform affordances enable divergent actors to co-produce collective narratives (Meraz & Papacharissi, 2013). The discourse’s orientation toward international institutions further highlights its transnational character: sanctions- and IRGC-focused streams show how diasporic actors reframe domestic grievances for global legitimacy, consistent with digital diaspora activism and transnational advocacy network theories (Keck & Sikkink, 2014; Sohrabi-Haghighat, 2011).

Semantic Coherence in a Transnational Discourse

Despite clear lexical separation across five streams, low Jensen–Shannon Divergence scores (0.06–0.15) and mixed MANOVA results indicate semantic proximity beyond simple fragmentation. Coherent clusters (modularity 0.61–0.73) form around expressive affinities rather than fixed ideologies (Bennett & Segerberg, 2012), producing a “polyvocal consensus” that contrasts with electoral polarization (Barberá et al., 2015) and supports differentiated information flows in digital movements (Jackson et al., 2020). Persistence of networked publics despite declining participation reflects sustained collective meaning-making rather than algorithmic artifact (Boyd, 2010; Olteanu et al., 2019). Transnational in orientation, the discourse enacts a “boomerang effect” through strategic public diplomacy (Keck & Sikkink, 1998), with sanctions and legal advocacy streams exemplifying “serial transnational activism” (Mercea & Bastos, 2016; Michaelsen, 2018). Methodologically, contrasting MANOVA results, thematic differentiation more visible in curated lexical space than raw embeddings—underscore the value of hybrid computational–interpretive approaches (Freelon & Wells, 2020; O’Connor et al., 2010).

Thematic Differentiation and Discourse Diversity

G² analysis reveals distinct stream vocabularies reflecting differentiated rhetorical strategies: Stream 1 (Documentation; e.g., “killed,” “bloody”) aligns with digital witnessing (Wulf et al., 2022), while Stream 4 (Monarchist Advocacy & Diaspora Political Engagement; e.g., “Reza,” “Pahlavi”) reflects diasporic memory work (Ghorashi & Boersma, 2009)—a “choreographed” diversity (Gerbaudo, 2012) within a single platform. Yet streams share a common semantic field: low JSD scores (0.06–0.15) indicate coherence consistent with “affective publics” (Papacharissi, 2015), and a t-test confirms intra-stream distances are significantly lower than inter-stream distances (M = 0.3430 vs. M = 0.4582, p = 0.0008). This “subjectivity in networks” (Milan, 2015) enables ideologically distinct streams to coexist. Rather than fragmentation, the pattern reflects strategic complementarity, streams fulfilling distinct functions from moral urgency to constituency mobilization, consistent with tactical repertoire diversification (Bennett & Segerberg, 2012; Keck & Sikkink, 1998). Statistical detection (topword MANOVA, p < 0.001) confirms structured organization, while boundary persistence supports “networked framing” (Meraz & Papacharissi, 2013).

These findings, however, likely reflect digitally empowered actors, educated English-speaking diaspora users, NGO professionals, and internationally connected activists rendering working-class Iranians, ethnic minorities (e.g., Kurds and Baluchis), and non-English speakers less visible. Stream 4’s near-neutral tone may reflect diasporic positional privilege relative to on-the-ground actors whose negativity is rooted in lived trauma. Overall, thematic diversity enhances movement effectiveness by broadening rhetorical reach while reducing vulnerability to information control (Michaelsen, 2018; Tufekci, 2017).

Affective Dimensions of Strategic Communication

VADER sentiment analysis reveals that each identified protest stream exhibits a distinct affective profile tightly aligned with its underlying communicative function, confirming that emotion operates as a strategically calibrated resource within transnational digital advocacy. Stream 3 (Campaigning against the IRGC and for International Sanctions) registers as the most intensely negative cluster (M = −0.257), a feature structurally consistent with denunciatory political discourse and the systematic production of moral outrage to capture international institutional attention (Papacharissi & de Fatima Oliveira, 2012). In sharp contrast, Stream 4 (Monarchist Advocacy and Diaspora Political Engagement) exhibits a near-neutral compound score (M = −0.028), reflecting a rhetorical pivot away from immediate trauma and toward institutional, forward-looking alternative governance models (Ghorashi & Boersma, 2009).

This systematic variation provides empirical support for the existence of strategically calibrated “affective publics” (Papacharissi, 2015), where emotional registers are intentionally tailored to match the expectations of different target audiences, such as state policymakers, foreign diplomats, or domestic activists. Crucially, this affective differentiation coexists with a shared baseline of negativity across the broader network, indicating a common grounding in collective grief, indignation, and urgency that functions as the foundational connective tissue for transnational mobilization (Bennett & Segerberg, 2012). These patterns point to the deployment of “strategic affect” rather than unmediated, spontaneous emotional outbursts (Gerbaudo, 2012; Milan, 2015). The temporal stability of these affective signatures across the 5-month sampling window suggests established rhetorical repertoires, demonstrating that emotional diversity can enhance collective action by distributing specialized emotional labor across semi-autonomous modules.

These findings align with prior computational literature investigating affectively structured political discourse on social media. For instance, Gündüz et al. (2024) demonstrate that targeted emotional registers enable structural cohesion within specialized, niche online communities. Similarly, Demirel and Gündüz (2022) establish that highly salient socio-political transformations generate structurally distinct sentiment communities characterized by polarized or highly differentiated emotional baselines (e.g., structural backlash vs. institutional affirmation). Within the #MahsaAmini English-language layer, Stream 3’s denunciatory outrage and Stream 4’s measured affective restraint operate not as polarizing cross-currents but as highly complementary rhetorical registers nesting within a unified civil society framework.

The concentration of neutral sentiment for Stream 4 (Monarchist Advocacy and Diaspora Political Engagement), which exhibits a near-neutral compound score ($M = −0.028$), this signature represents more than an absence of emotion; it indexes three analytically distinct conditions: strategic institutionalism, positional ambivalence, and reform fatigue.

First, neutrality functions as an active rhetorical tool for strategic de-escalation and institutional legitimacy. Transnational networks targeting institutional gatekeepers, such as foreign ministries and international NGOs, must deliberately suppress high-arousal emotional registers to match the technocratic, formal, and legalistic expectations of international diplomacy. For Stream 4, maintaining a neutral affective signature serves to foster passive trust among international observers, projecting legal stability and governance preparedness. Second, neutral sentiment reflects a structural positional and diasporic ambivalence. As the monarchist stream balances factional ideology against the broader #MahsaAmini umbrella, its communicative output relies on descriptive, factual assertions that create a “neutrality of friction.” Third, the prevalence of neutral sentiment serves as a linguistic proxy for reform fatigue and emotional exhaustion. This neutral baseline represents a strategic pivot toward long-term organizational endurance, marking an adaptation where emotional expressions are rationalized into structural documentation and policy reporting.

Within the overarching modular polyvocality framework, these diverse expressions of neutrality are not analytical contradictions but vital functional complements. While Streams 1 and 3 supply the network’s core moral urgency and disruptive outrage, Stream 4’s affective restraint broadens the protests overall rhetorical and strategic range. This distributed affective labor enables the network to simultaneously energize emotionally mobilized activists on the ground while addressing formal, institutional audiences abroad—a structural synergy that reinforces collective persistence across informationally constrained environments without requiring uniform emotional expression.

English as Platform Strategy

English functioned as a strategic communicative layer with its own rhetorical logic. The discourse enacts a “boomerang effect” through international political vocabulary generating external pressure and strategic public diplomacy (Keck & Sikkink, 1998), aligning with “serial transnational activism” that reframes local struggles for global policymakers (Mercea & Bastos, 2016; Michaelsen, 2018). English also enabled platform optimization by leveraging “networked gatekeeping” for algorithmic visibility and global reach (Meraz & Papacharissi, 2013; Shoemaker & Vos, 2009), while serving as a coordination tool for diaspora groups and NGOs (Zajak, 2014). Language choice thus operates as strategic resource allocation enhancing movement effectiveness through differentiated communicative targeting.

Temporal Evolution and Strategic Adaptation

A chi-square test between Period 1 (Oct–Dec 2022) and Period 2 (Jan–Feb 2023) confirms a significant discursive shift (p < 0.001): from local mobilization and victim testimony toward international advocacy and sanctions framing, with legal and geopolitical terms increasing as mourning vocabulary declined. This reflects strategic adaptation to intensified state repression, shifting discourse from testimonial narration to external pressure strategies. Hierarchical clustering reveals two meta-streams corresponding to these temporal phases, indicating that time actively structures discourse formation. Thematic boundaries function as evolving repertoires shaped by political urgency and audience targeting, consistent with connective action theory’s account of digital publics recalibrating narratives under shifting constraints (Bennett & Segerberg, 2012).

Networked Gatekeeping and Transnational Framing

Visibility in the #MahsaAmini discourse was shaped by networked rather than centralized gatekeeping, operating through engagement patterns, hashtags, and algorithmic amplification (Meraz & Papacharissi, 2013; Shoemaker & Vos, 2009). Each stream functioned as a semi-independent gatekeeping node: monarchist content leveraged influencers and nationalist hashtags; documentation streams circulated witness imagery; sanction-related streams targeted policymakers and journalists. Despite lexical overlap, framing logics diverged from mourning to diplomatic urgency, reflecting a decentralized editorial process in which distributed actors collectively structure global visibility through multiple semi-autonomous framing nodes.

Limitations and Future Research

Several limitations warrant acknowledgment. First, the English-language focus excludes Persian-language discourse, biasing findings toward globally engaged actors and likely underrepresenting local resistance; without granular user data, diaspora and domestic participation cannot be distinguished. Future work should extend to cross-linguistic and cross-platform comparisons (e.g., Instagram and Telegram). Second, qualitative coding analyzed 75 tweets per community, risking underrepresentation of rare themes (<10%), within-community heterogeneity, and elite voice overrepresentation. These concerns are mitigated by triangulation across G² analysis, Jensen–Shannon Divergence, MANOVA, and sentiment analysis, whose convergence with qualitative results suggests robust patterns; future work could apply large-scale topic modeling to capture sub-stream structures. Third, stream labeling retains interpretive subjectivity despite longitudinal definitional consistency. Fourth, bot detection remains heuristic rather than ensemble-based (e.g., Botometer), warranting more advanced frameworks. Fifth, network analysis captures retweet and co-hashtag alignment but not dialogic interaction (replies, mentions), and temporal dynamics could benefit from automated frame detection. Finally, future research should link discourse structures to offline outcomes, protest activity, and policy impact through multi-lingual, multi-modal, and multi-scalar designs.

While our analysis identifies modular thematic coherence, we acknowledge that this may be influenced by structural factors beyond organic user intent. Future research should disentangle the role of algorithmic amplification where platform recommendation systems push ideologically consistent content and coordinated campaigning where organized actors may be artificially inflating the perceived unity of these streams. Furthermore, our findings are constrained by an English-language sampling bias, which may overlook critical vernacular nuances present in non-English protest discourse.

Conclusion

By introducing and empirically validating modular polyvocality, this study demonstrates that transnational protest movements can maintain multiple thematic streams while preserving a shared affective and rhetorical core. Our integration of Louvain community detection, G² keyword profiling, semantic divergence measures, and intercoder-verified qualitative coding offers a methodological template for analyzing hybrid cohesion in large-scale online discourse. Applied to the English-language #MahsaAmini mobilization, this framework reveals how activists strategically recalibrated their discourse from local testimony to international advocacy by maintaining the convergence of the movement.

The Modular Polyvocality Framework: A Synthesis

The concept of modular polyvocality offers a middle-range theory for understanding how contemporary movements sustain convergence and cohesion across diversity. It synthesizes insights from three literatures: (1) network science, which provides structural foundation (modularity as community clustering), (2) discourse analysis, which contributes rhetorical dimension (polyvocality as voice multiplicity), and (3) social movement studies, which supplies functional logic (strategic complementarity as distributed labor). Where existing frameworks emphasize either coordination mechanisms (connective action) or affective dynamics (affective publics), modular polyvocality specifies the structural outcome: discourse architecture characterized by thematic specialization within affective alignment. This pattern appears particularly adaptive for transnational movements operating under three conditions: (a) normative framing permitting tactical flexibility, (b) multiple target audiences requiring differentiated messaging, and (c) information environments where redundancy and diversification enhance resilience. Future research should test whether modular polyvocality generalizes beyond the Iranian case to other movements with similar socio-political roots, from Hong Kong’s pro-democracy protests to climate activism’s international coordination.

Key Empirical Contributions

We found the discourse exhibited modular polyvocality: thematic differentiation within broader rhetorical coherence. The identification of five persistent and complementary streams with distinct lexical and affective signatures indicates systematic organization:

1. Documentation of events involving violence.

2. On-the-Ground Mobilization.

3. Anti-IRGC Advocacy and Sanctions Campaigns.

4. Monarchist Advocacy & Diaspora Political Engagement.

5. Broader Opposition Narratives & International Lobbying.

Our temporal analysis revealed strategic adaptation from on-the-ground event terms to international advocacy language, showing how social mobilizations recalibrate in response to changing political conditions.

Theoretical Implications

These findings contribute to three scholarly conversations. They support connective action theories by showing how mobilizations maintain coherence without centralized coordination, enabling coalition politics based on strategic complementarity (Bennett & Segerberg, 2012). They extend networked publics concepts by demonstrating how strategic communication choices interact with platform affordances to create sustained thematic organization (Boyd, 2010; Papacharissi, 2015). Finally, they advance transnational activism literature by revealing how English serves strategic functions beyond simple translation, helping mobilizations develop specialized competencies for global engagement.

Methodological Insights

Our multi-method approach demonstrates computational discourse analysis potential and constraints. The contrasting MANOVA results (no separation in full embedding space vs. significant separation in top-word space) highlight the conditional nature of thematic detectability, underscoring the value of our integrated framework, which maintains analytical rigor while balancing computational scale with qualitative depth.

Final Reflections

We propose the modular polyvocality framework has utility beyond the #MahsaAmini case. Protest operating in restrictive political contexts, multilingual publics, and transnational advocacy demands may rely on similar patterns of thematic differentiation within affective coherence. Recognizing and measuring this structure can help scholars and practitioners understand how cohesion is sustained without homogeneity and how digital publics adapt rhetorical repertoires over time. By foregrounding this framework, we invite further comparative and cross-linguistic applications that can refine our understanding of cohesion, diversity, and strategic adaptation in the evolving grammar of political resistance.

Supplemental Material

Supplemental material - Modular Polyvocality: Community Detection and Semantic Analysis of Civil Society Discourse in English-Language #MahsaAmini Twitter Networks

Supplemental material for Modular Polyvocality: Community Detection and Semantic Analysis of Civil Society Discourse in English-Language #MahsaAmini Twitter Networks by Hossein Masoudnia, Alireza Samiee Esfahani, Mohsen Forghani, Mohamamd Moghimi, and Hossein Samiei Esfahany in Social Science Computer Review

Footnotes

ORCID iD

Mohsen Forghani

Ethical Considerations

This study analyzes publicly available Twitter data on #MahsaAmini (October 2022–February 2023). We recognize the heightened ethical risks in politically sensitive contexts under extreme surveillance, where user identification could expose individuals to harm (Kermani, 2025; Wulf et al., 2022). Accordingly, the ethical framework prioritizes user protection. First, only publicly available tweets accessed via Twitter’s public API were used; no private or direct-message data was collected. Second, all results are reported at aggregate levels, with no individual tweets, usernames, or behavioral traces disclosed in any form that could enable identification. Third, all user identifiers and metadata were removed prior to analysis, and communities are reported using analytical labels only. Network visualizations exclude readable node identifiers. The dataset used (Banachewicz, 2023; CC0 license, Kaggle) was accessed in compliance with platform policies at the time of collection. All Supplemental Material used for validation was fully anonymized. Following established guidelines for computational social media research (Franzke et al., 2020; Townsend & Wallace, 2017), the study did not require formal IRB approval due to its aggregate-only design and anonymization procedures. Nevertheless, we acknowledge ongoing ethical responsibilities in sensitive political contexts and applied precautionary safeguards accordingly. This study analyzed publicly available Twitter data related to the #MahsaAmini protests (October 2022–February 2023). No private or direct-message content was accessed, and no attempts were made to contact individual users. Usernames and identifying details were removed or anonymized in all datasets, tables, and visualizations. Data collection complied with Twitter’s Developer Policy and relevant terms of service at the time of access. In line with established ethical guidelines for research using public social media data (Franzke et al., 2020; Townsend & Wallace, 2017), this study did not require institutional review board (IRB) approval.

Consent to Participate

Not applicable, as no human participants as research subjects in this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data Availability Statement

The raw Twitter data analyzed in this study was obtained from the IranProtest Tweets dataset by Konrad Banachewicz, available on Kaggle under a CC0: Public Domain license (https://www.kaggle.com/datasets/konradb/iranprotest-tweets). The original dataset contains tweets from September 23, 2022, to June 11, 2023, and is publicly accessible for research purposes. Supplementary materials are available on Figshare (), with an additional anonymized CSV of sample tweets for labeling validation provided via the journal’s submission system. Supplementary materials are available on Figshare with an additional anonymized CSV of sample tweets for labeling validation provided via the journal’s submission system. The computational code used for data processing and analysis is available upon reasonable request from the corresponding author. All analytical procedures are fully documented in the methodology section to ensure reproducibility.

Supplemental Material

Supplemental material for this article is available online.

Author Biographies

Hossein Masoudnia is Associate Professor of Political Science at the Faculty of Administrative Sciences and Economics, University of Isfahan, Iran. His research focuses on political sociology, social media and political behavior, populism, and comparative politics in the Middle East. He has published widely on topics including the relationship between social media consumption and political distrust, rentier state theory, and governance quality under populist leadership.

Alireza Samiee Esfahani is Associate Professor of Political Science at the Faculty of Administrative Sciences and Economics, University of Isfahan, Iran. His research specializations include political theories, methodology in political science, comparative politics, and Middle East studies, with particular focus on cultural globalization, Islamic politics, and Iranian foreign policy. He has published across Persian and English-language political science journals.

Mohsen Forghani is a Geospatial Data Analyst and Urban Planning Specialist at Sari Municipality, Mazandaran, Iran. He holds an MA in Geography and Urban Planning from the University of Mazandaran. His research sits at the intersection of computational social science and urban informatics, with expertise in network analysis, natural language processing, and spatiotemporal methods applied to politically sensitive social contexts.

Mohammad Moghimi is a PhD Researcher at the SFI Centre for Research Training in Artificial Intelligence (CRT-AI), University of Limerick, Ireland, supervised by Dr. Patrick Healy. His doctoral research focuses on Sound Event Detection for autonomous vehicles using neural networks and deep learning. He holds an MSc in Electronic Engineering with a specialization in sound and music processing, and has professional experience in signal processing research and software development.

Hossein Samiee Esfahani holds an MA in Geopolitics from Ferdowsi University, Iran. His research focuses on the structural analysis of political institutions in Middle Eastern countries, with particular interest in the geopolitical dynamics shaping state formation and institutional change across the region.

Appendix

Appendix A:

Thematic Coding Framework for Community-Level Discourse Analysis

Stream title	Core definition & purpose	Primary indicators & content focus	Characteristic keywords & hashtags
1. Documentation of events involving violence & victim advocacy	To bear witness to the human cost by documenting gender-related harms, naming victims, and publicizing atrocities.	- Naming victims: Focus on individual stories (Mahsa, Nika, Sarina).- documenting violence: Details of torture, rape, killing.- gendered violence: Schoolgirl poisonings, morality police actions.- Memorialization: Marking chehelom/anniversaries.	Killed, torture, prison, victim, shot, schoolgirls, beaten, saytheirnames, names of victims
2. On-the-ground mobilization & protest narratives	To chronicle the revolution in action by reporting on the logistics, location, and spirit of protests, strikes, and defiance.	- Reporting Street protests: Naming specific cities, universities.- collective action: Worker/student strikes, business closures.- acts of defiance: Hijab burning, anti-regime chanting.- general revolutionary calls: Broad, non-factional slogans.	Protests, strike, streets, revolution, freedom, chanting, university, names of cities
3. Campaigning against IRGC & for international sanctions	To conduct a targeted external campaign pressuring international bodies to take punitive action, specifically against the IRGC.	- IRGC designation calls: The central demand for sanction listing.- targeting officials: Directly addressing foreign politicians (borrell, etc.).- linking to global threats: Framing the IRGC as an international security risk (e.g., Ukraine).	irgc, sanctions, blacklist, expel, #IRGC, eu, official handles
4. Monarchist advocacy & diaspora political engagement	To promote the restoration of the pahlavi monarchy and position reza pahlavi as the legitimate leader of the opposition.	- Promoting pahlavi: Positive mentions, images, tags of reza pahlavi.- monarchist hashtags: Use of #JavidShah, etc.- critique of other factions: Negative framing of MEK, reformists.- pre-revolutionary Nostalgia.	Pahlavi, reza, shah, javidshah, kingdom, prince, crown
5. Broader opposition narratives & lobbying	To represent the discourse of other organized factions (e.g., MEK/NCRI) and thematic campaigns against perceived regime allies (e.g., NIAC).	- Support for MEK/NCRI: Promoting Maryam rajavi, “resistance units”- Anti-NIAC sentiment: Aggressive criticism of NIAC as “lobbyists”- rejection of monarchy: “No to shah, no to leader” framing.- specific external campaigns: Organizing around events (FIFA, WalkoutIRI).	Mek, ncri, rajavi, niac, lobbyist, resistance, walkoutiri, parliament

References

Ahmed

Jaidka

(2013). Protests against #delhigangrape on Twitter: Analyzing India's Arab spring. eJournal of eDemocracy and Open Government, 5(1), 28–58.

Akbari

(2022). Iran: Digital spaces of protest and control piloting report on the guide on digitally-mediated assemblies. European center for not-for-profit law stichting (ECNL). European Center for non profit law.

Al-azawi

Safaa

(2022). Feature extractions and selection of bot detection on Twitter: A systematic literature review. Inteligencia Artificial, 25(69), 57–86. https://doi.org/10.4114/intartif.vol25iss69pp57-86

Atefinia

Mirnezami

S. R.

(2025). Exploring Iranian sentiments on the paris agreement: Insights from Twitter. Heliyon.

Banachewicz

(2023). Iran protests - Tweets, version 272. Kaggle. https://www.kaggle.com/datasets/konradb/iran-protests-2022-tweets

Barberá

Jost

J. T.

Nagler

Tucker

J. A.

Bonneau

(2015). Tweeting from left to right: Is online political communication more than an echo chamber? Psychological Science, 26(10), 1531–1542. https://doi.org/10.1177/0956797615594620

Bennett

W. L.

Segerberg

(2012). The logic of connective action: Digital media and the personalization of contentious politics. Information, Communication & society, 15(5), 739–768.

Bird

Klein

Loper

(2009). Natural language processing with python: Analyzing text with the natural language toolkit. O'Reilly Media, Inc.

Boyd

(2010). Streams of content, limited attention: The flow of information through social media. EDUCAUSE Review.

10.

Braun

Clarke

(2006). Using thematic analysis in psychology. Qualitative Research in Psychology, 3(2), 77–101. https://doi.org/10.1191/1478088706qp063oa

11.

Campbell

W. M.

Dagli

C. K.

Weinstein

C. J.

(2013). Social network analysis with content and graphs. Lincoln Laboratory Journal, 20(1), 61–81.

12.

Castillo

(2016). Big crisis data: Social media in disasters and time-critical situations. Cambridge University Press.

13.

Castillo

Mendoza

Poblete

(2011). Information credibility on Twitter. In Proceedings of the 20th international conference on world wide web. (pp. 675–684). Association for Computing Machinery.

14.

Davis

C. A.

Varol

Ferrara

Flammini

Menczer

(2016). BotOrNot: A system to evaluate social bots. In Proceedings of the 25th international conference companion on world wide web (pp. 273–274). Association for Computing Machinery.

15.

Demirel

Gündüz

(2022). World health organization's Twitter use before and during Covid-19 pandemic. World Health Organization.

16.

Demirel

Kahraman Gökalp

Gündüz

(2024). A text mining analysis of the change in status of the Hagia sophia on Twitter: The political discourse and its reflections on the public opinion. Atlantic Journal of Communication, 32(1), 63–90. https://doi.org/10.1080/15456870.2022.2093354

17.

Dunning

T. E.

(1993). Accurate methods for the statistics of surprise and coincidence. Computational Linguistics, 19(1), 61–74.

18.

Etemadifard

S. M.

Khazaei

(2021). Divergent heterogeneities in Iranian migration in France: Semantic and theoretical limitations of Diaspora. International Migration, 59(4), 207–220. https://doi.org/10.1111/imig.12779

19.

Farzam

Moradi

Mohammadi

Padar

Siegel

A. A.

(2023). Opinion manipulation on farsi Twitter. Scientific Reports, 13(1), 333. https://doi.org/10.1038/s41598-022-26921-5

20.

Franzke

A. S.

Bechmann

Ess

, et al. (2020). Internet Research: Ethical Guidelines 3.0. AoIR (The International Association of Internet Researchers.

21.

Freelon

Wells

(2020). Disinformation as political communication. Political communication.

22.

Gerbaudo

(2012). Tweets and the streets: Social media and contemporary activism. Pluto Press.

23.

Ghorashi

Boersma

(2009). The 'Iranian diaspora' and the new media: From political action to humanitarian help. Development and Change, 40(4), 667–691. https://doi.org/10.1111/j.1467-7660.2009.01567.x

24.

Golkar

(2015). Student activism, social media, and authoritarian rule in Iran. In The whole world is texting: Youth protest in the information age (pp. 61–79). Sense Publishers.

25.

Guest

Bunce

Johnson

(2006). How many interviews are enough? An experiment with data saturation and variability. Field Methods, 18(1), 59–82. https://doi.org/10.1177/1525822X05279903

26.

Gündüz

Demirel

Tombul

(2024). Exploring the concept of financial domination on social media: Sentiment and text analysis on Twitter. Atlantic Journal of Communication, 32(4), 602–625. https://doi.org/10.1080/15456870.2023.2178000

27.

Haralabopoulos

Simperl

(2017). Crowdsourcing for beyond polarity sentiment analysis: A pure emotion lexicon. arXiv preprint arXiv:1710.04203.

28.

Hutto

Gilbert

(2014). VADER: A parsimonious rule-based model for sentiment analysis of social media text. Proceedings of the International AAAI Conference on Web and Social Media, 8(1), 216–225. https://doi.org/10.1609/icwsm.v8i1.14550

29.

Jackson

Stromer-Galley

Hemsley

(2020). Differentiated information flows: Social media curation practices in the 2016 US presidential elections. International Journal of Communication, 14(1), 22.

30.

Keck

M. E.

Sikkink

(1998). Transnational advocacy networks in the movement society. The social Movement society: Contentious Politics for a New Century, 221(2), 217–237.

31.

Keck

M. E.

Sikkink

K. A.

(2014). Activists beyond borders: Advocacy networks in international politics. Cornell university press.

32.

Keller

F. B.

Klinger

(2019). Social bots in election campaigns: Theoretical, empirical, and methodological implications. Political Communication, 36(1), 171–189. https://doi.org/10.1080/10584609.2018.1526238

33.

Kermani

(2023). Profile: #MahsaAmini: Iranian Twitter activism in times of computational propaganda. Social Movement Studies, 24(2), 257–267. https://doi.org/10.1080/14742837.2023.2180354

34.

Kermani

(2025). The art of delirium: Social media suppression in authoritarian regimes. Communication Theory, 35(4), qtaf006.

35.

Khatam

(2023). Mahsa Amini's killing, state violence, and moral policing in Iran. Human Geography, 16(3), 299–306. https://doi.org/10.1177/19427786231159357

36.

Khorramrouz

Dutta

KhudaBukhsh

A. R.

(2023). For women, life, freedom: A participatory AI-based social web analysis of a watershed moment in Iran's gender struggles. arXiv Preprint arXiv:2307.03764.

37.

Kosmajac

Keselj

(2019). Twitter bot detection using diversity measures. In Proceedings of the 3rd international conference on natural language and speech processing (pp. 1–8). Association for Computational Linguistics.

38.

Krippendorff

(2013). Content analysis: An introduction to its methodology. Sage Publications.

39.

Meraz

(2017). Hashtag wars and networked framing: The private/public networked protest repertoires of occupy on Twitter. In Between the public and private in mobile communication (pp. 303–323). Routledge.

40.

Meraz

Papacharissi

(2013). Networked gatekeeping and networked framing on #egypt. The International Journal of Press/Politics, 18(2), 138–166. https://doi.org/10.1177/1940161212474472

41.

Mercea

Bastos

M. T.

(2016). Being a serial transnational activist. Journal of Computer-Mediated Communication, 21(2), 140–155. https://doi.org/10.1111/jcc4.12150

42.

Michaelsen

(2018). Exit and voice in a digital age: Iran’s exiled activists and the authoritarian state. Globalizations, 15(2), 248–264. https://doi.org/10.1080/14747731.2016.1263078

43.

Michelen

Glasius

(2018). Authoritarian practices in the digital age—Introduction. International Journal of Communication.

44.

Milan

(2015). From social movements to cloud protesting: The evolution of collective identity. Information, Communication & Society, 18(8), 887–900. https://doi.org/10.1080/1369118x.2015.1043135

45.

Mohammadi

Moradi

Firouzabadi

S. M.

Jafari

(2022). The footprint of campaign strategies in farsi Twitter: A case for 2021 Iranian presidential election. PLOS ONE, 17(7), Article e0270822. https://doi.org/10.1371/journal.pone.0270822

46.

O'Connor

Balasubramanyan

Routledge

Smith

(2010). From tweets to polls: Linking text sentiment to public opinion time series. Proceedings of the International AAAI Conference on Web and Social Media, 4(1), 122–129. https://doi.org/10.1609/icwsm.v4i1.14031

47.

Olteanu

Castillo

Diaz

Kıcıman

(2019). Social data: Biases, methodological pitfalls, and ethical boundaries. Frontiers in Big Data, 2(1), 13. https://doi.org/10.3389/fdata.2019.00013

48.

Pacheco

Hui

P. M.

Torres-Lugo

Truong

B. T.

Flammini

Menczer

(2021). Uncovering coordinated networks on social media: Methods and case studies. Proceedings of the International AAAI Conference on Web and social Media, 15(1), 455–466. https://doi.org/10.1609/icwsm.v15i1.18075

49.

Papacharissi

(2015). Affective publics: Sentiment, technology, and politics. Oxford University Press.

50.

Papacharissi

(2016). Affective publics and structures of storytelling: Sentiment, events and mediality. Information, Communication & society, 19(3), 307–324. https://doi.org/10.1080/1369118x.2015.1109697

51.

Papacharissi

de Fatima Oliveira

(2012). Affective news and networked publics: The rhythms of news storytelling on #egypt. Journal of Communication, 62(2), 266–282. https://doi.org/10.1111/j.1460-2466.2012.01630.x

52.

Pavan

Rapini

(2022). Battles over issues in networked publics: Investigating the discursive mobilization of the antifascist frame on Twitter. Mobilization, 27(1), 69–90. https://doi.org/10.17813/1086-671x-27-1-69

53.

Piper

Uhlin

(2004). Transnational activism in Asia. Problems of power and democracy. Routledge.

54.

Rauchfleisch

Kaiser

(2020). The false positive problem of automatic bot detection in social science research. PloS One, 15(10), Article e0241045. https://doi.org/10.1371/journal.pone.0241045

55.

Řehůřek

Sojka

(2010). Software Framework for Topic Modelling with Large Corpora. Proceedings of LREC 2010 workshop New Challenges for NLP Frameworks.

56.

Rosenbaum

J. E.

Bouvier

(2020). Twitter, social movements and the logic of connective action: Activism in the 21st century–an introduction. Participation: Journal of Audience & Reception Studies, 17(1), 120–125.

57.

Shoemaker

P. J.

Vos

(2009). Gatekeeping theory. Routledge.

58.

Sohrabi-Haghighat

M. H.

(2011). New media and social-political change in Iran. CyberOrient, 5(1), 90–109. https://doi.org/10.1002/j.cyo2.20110501.0004

59.

Stewart

Schultze

(2019). Producing solidarity in social media activism: The case of my stealthy freedom. Information and Organization, 29(3), Article 100251. https://doi.org/10.1016/j.infoandorg.2019.04.003

60.

Sunstein

C. R.

(2018). Republic: Divided democracy in the age of social media. Princeton University Press.

61.

Tabachnick

B. G.

Fidell

L. S.

(2013). Using multivariate statistics (6. Baskı). MA: Pearson.

62.

Tan

D’Souza

(2025). Bridging the evaluation gap: Leveraging large language models for topic model evaluation. 21st Conference on. Information and Research science Connecting to Digital and Library science.

63.

Townsend

Wallace

(2017). The ethics of using social media data in research: A new framework. The ethics of online research.

64.

Tufekci

(2017). Twitter and tear gas: The power and fragility of networked protest. Yale University Press.

65.

Varol

Ferrara

Davis

Menczer

Flammini

(2017). Online human-bot interactions: Detection, estimation, and characterization. Proceedings of the International AAAI Conference on Web and Social Media, 11(1), 280–289. https://doi.org/10.1609/icwsm.v11i1.14871

66.

Wulf

Randall

Aal

Rohde

(2022). The personal is the political: Internet filtering and counter appropriation in the Islamic Republic of Iran. Computer Supported Cooperative Work, 31(2), 373–409. https://doi.org/10.1007/s10606-022-09426-7

67.

Zajak

(2014). Pathways of transnational activism: A conceptual framework (No. 14/5). MPIfG Discussion Paper.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.04 MB

Modular Polyvocality: Community Detection and Semantic Analysis of Civil Society Discourse in English-Language #MahsaAmini Twitter Networks

Abstract

Keywords

Introduction

Conceptualizing Modular Polyvocality

Literature Review

Transnational Activism and the Strategic Use of English

The Iranian Twittersphere as a Contested Transnational Arena

Theoretical Framework: Networked Publics, Gatekeeping, and Framing

From Fragmentation to Modular Polyvocality: Theorizing Discourse Structure

Our Contribution

Cross-Cutting Analyses

Materials and Methods

Data Collection and Preprocessing

Community Detection Algorithm

Bot Flagging

Stream Mapping and Reliability Assessment

Reliability Assessment

Lexical Characterization and Topic Divergence

Lexical Characterization (G2 Keyword Profiling)

Topic Divergence (Jensen–Shannon Divergence)

Semantic Validation

Word Embedding Approach

Semantic Coherence Tests

Semantic Space Mapping via Multidimensional Scaling (MDS)

Statistical Validation (MANOVA Tests)

Temporal and Hierarchical Analyses

Hierarchical Clustering: Meta-Streams

Semantic Drift: Month-to-Month Lexical Change

Sentiment Analysis Using VADER

Results

Network Modularity and Community Structure

Thematic Stream Reliability Validation

The Contested Information Environment: Bot-Like Activity Across Streams

Lexical Distinctions Across Streams

Topic Divergence (JSD Matrix)

Stream Validation—Theme–Centroid Alignments

Meta-Streams and Temporal Drift

MANOVA Findings: Statistical Differentiation of Streams

Stream Evolution Over Time

Affective Signatures: Sentiment Analysis of Discourse Streams

Discussion: Modular Polyvocality in Transnational Digital Activism

Semantic Coherence in a Transnational Discourse

Thematic Differentiation and Discourse Diversity

Affective Dimensions of Strategic Communication

English as Platform Strategy

Temporal Evolution and Strategic Adaptation

Networked Gatekeeping and Transnational Framing

Limitations and Future Research

Conclusion

The Modular Polyvocality Framework: A Synthesis

Key Empirical Contributions

Theoretical Implications

Methodological Insights

Final Reflections

Supplemental Material

Supplemental material - Modular Polyvocality: Community Detection and Semantic Analysis of Civil Society Discourse in English-Language #MahsaAmini Twitter Networks

Footnotes

ORCID iD

Ethical Considerations

Consent to Participate

Funding

Declaration of Conflicting Interests

Data Availability Statement

Supplemental Material

Author Biographies

Appendix

References

Supplementary Material

Lexical Characterization (G² Keyword Profiling)