Telegram Study Reveals Selective Topical Cohesion Inside Far-Right Forwarding Networks

A computational network investigation examining communication patterns across far-right Telegram channels demonstrates that message diffusion is governed by selective topical cohesion rather than uniform ideological dissemination. Analyzing thousands of forwarded posts, the study reveals that core ideological narratives such as white nationalism form insular, tightly knit sharing clusters, whereas cross-cutting themes such as immigration circulate diffusely across otherwise separated communities. Conducted by researcher Yasmine Houri and published in Applied Network Science, the research addresses the long-standing problem of opinion homophily in digital communication environments. By combining retrieval-augmented language modeling with network science metrics, the investigation establishes how specific discursive subjects systematically dictate the structural architecture of message sharing. The findings provide quantitative evidence that polarization within alternative messaging spaces does not emerge solely through broad ideological alignment, but develops through the precise interplay between particular thematic discussions and directed forwarding pathways.

The study examines Telegram because the messaging service has functioned as a central destination for controversial online actors following large-scale deplatforming interventions on mainstream social media. Stricter moderation policies and extensive account removals across major public platforms disproportionately affected groups associated with white nationalism, antisemitism, and the alt-right, prompting many public figures to announce public migrations to alternative digital environments. Unlike conventional feed-based social media that rely on algorithmic content recommendations, Telegram provides cloud-based broadcast channels and group settings where creators and administrators can disseminate material anonymously to subscribers. The platform incorporates specific affordances designed around social privacy, which allows participants to minimize personal disclosures through features including anonymous forwarding, message editing, and post removal. These technical characteristics attract fringe communities seeking to evade content moderation while maintaining audience reach, establishing an ecosystem where message forwarding serves as the primary mechanism for cross-channel content circulation and community interconnection.

To investigate these communication dynamics systematically, the researcher analyzed historical data extracted from the Pushshift Telegram Dataset, an open-source repository containing over three hundred million messages from nearly twenty-eight thousand channels collected between September 2015 and November 2019. Focusing on this pre-pandemic baseline, the study identified relevant far-right communities using a curated list of ideological access terms established in prior academic literature. After matching channel titles, usernames, and descriptions against these markers, the initial collection underwent four rigorous filtering stages: retaining only English-language posts identified through automated transformer classification, excluding inactive accounts with fewer than five relevant posts, removing off-topic content through manual inspection, and eliminating channels dominated by image captions or spam. This curation yielded a focused corpus of 61,346 messages published across 66 distinct channels between October 1, 2015, and October 22, 2019. Longitudinal analysis revealed that message frequency surged dramatically during 2019, reflecting the rapid arrival of newly relocated communities, while total posting volume followed a heavily skewed Pareto distribution dominated by a handful of prolific accounts such as brexiteerschatlounge.

Capturing the topical complexity of informal and coded political rhetoric required developing an automated multi-label classification pipeline capable of handling overlapping discursive themes without imposing mutually exclusive categories. The author created a retrieval-augmented few-shot classification framework that combined dense vector embeddings with generative language processing. To define category boundaries, a stratified sample of 230 messages drawn across channel-month pairs was manually annotated through iterative coding cycles, establishing eleven distinct topic labels alongside formal definitions. Automated processing deployed the mistral-embed model to project messages into a 1024-dimensional semantic space, followed by cosine similarity searches to retrieve the nearest annotated exemplar for each unclassified text. The generation model, mistral-large-2411 operating within a beta conversation endpoint, received the dynamic contextual example alongside explicit prompt instructions. To guarantee deterministic output and classification reproducibility, generation parameters were fixed at zero temperature, top-p equal to one, and a maximum token allocation of 2048, bypassing conventional tokenization preprocessing while preserving multi-label integrity across varied message lengths.

The retrieval-augmented framework was benchmarked against several standard natural language processing baselines, including a fine-tuned BERT classifier, zero-shot natural language inference via Facebook’s BART model executed on an NVIDIA GeForce RTX 3080 Laptop GPU, zero-shot generation using Gemini 2.5 Flash Lite, and Mistral without retrieval augmentation. Across comparative evaluation metrics encompassing precision, recall, micro-averaged F1 score, and subset accuracy, the retrieval-augmented Mistral configuration demonstrated superior performance, attaining a subset accuracy of 0.743 despite the strict mathematical penalization inherent to multi-label exact matching. The architecture achieved robust precision while sustaining the elevated recall essential for downstream network modeling, ensuring that secondary themes within complex posts were not overlooked. Furthermore, the retrieval framework delivered these outcomes using only 230 annotated exemplars without requiring parameter fine-tuning, providing a computationally efficient and methodologically transparent alternative to resource-intensive transformer training workflows when categorizing noisy social media corpora.

Descriptive evaluation of the classified corpus showed that approximately 35 percent of all posts fell under the catch-all category of None of the above, illustrating that curated political channels frequently publish material outside predefined ideological boundaries. Among the categorized discourse, the five most prominent subjects were Online coordination at 13.3 percent, Anti-establishment, lies and disinformation at 8.9 percent, UK politics and Brexit at 8.1 percent, Nationalism and white supremacy at 7.2 percent, and Proud Boys/Antifa and the manosphere at 7.0 percent. The prominence of operational planning indicated that collective organizing represents a primary structural function within these channels. Pairwise co-occurrence calculations revealed that Online coordination appeared in four of the five most frequent topic pairings, often accompanying Brexit discussions. Moreover, while eight of the ten topical categories exhibited co-occurrence rates exceeding 68 percent, themes like Nationalism and white supremacy appeared in isolation substantially more often, pointing toward their specialized role as core narrative pillars that are nevertheless discussed across multiple distinct channels.

To model information circulation, the author constructed a directed, weighted network using 21,273 forwarded messages identified within the corpus, where nodes represented channels and directed edges denoted the forwarding of material from origin to recipient. Edge weights were normalized by each recipient channel’s total volume of incoming forwards to emphasize relative prominence rather than raw message counts. To preserve thematic coherence and address the classical boundary specification problem in network science, the analysis excluded out-of-sample source channels, focusing exclusively on the 45 curated channels that actively shared content among one another. Divergence metrics confirmed that this reduced sample maintained near-perfect distributional equivalence with the broader 66-channel corpus. The resulting forwarding graph comprised 45 nodes and 145 directed edges, containing a weakly connected component of 43 nodes and a strongly connected core of only 13 channels. Modularity clustering using the Louvain algorithm identified four descriptive ideological groupings corresponding to Brexit proponents, the Proud Boys, white supremacists, and fascist entities, while heavy-tailed degree distributions indicated that content dissemination was driven by a limited number of central distribution hubs.

Channel-level information diversity was quantified by calculating the Shannon entropy of each channel’s topic prevalence vector, which measured content dispersion and predictability across messages. Observed entropy scores exhibited a left-skewed distribution in which the median value of 2.56 bits exceeded the mean of 2.10 bits, demonstrating that while a small contingent of channels maintained an exceptionally narrow and repetitive thematic focus, the majority of channels engaged with a relatively wide spectrum of subjects. When examining the relationship between channel activity and narrative breadth, the analysis found only a weak positive correlation of r = 0.28 between total message volume and Shannon entropy. This loose association confirms that high posting volume does not automatically produce diverse discourse, revealing instead that far-right actors deploy contrasting editorial strategies, ranging from highly specialized messaging designed to reinforce core doctrines to generalist publishing intended to capture broader audiences.

The central structural findings emerged from calculating Newman’s assortativity coefficients across individual topic prevalence vectors on the forwarding graph, paired with non-parametric permutation significance tests. Eight of the ten evaluated topics exhibited statistically significant positive assortativity at p <= 0.01, demonstrating that message forwarding is structured along clear lines of topical alignment rather than uniform mixing. The strongest assortative clustering appeared in foundational identity themes, led by Nationalism and white supremacy with a coefficient of r = 0.71, followed by Proud Boys/Antifa and the manosphere at r = 0.62, and Anti-left rhetoric at r = 0.58. Conversely, cross-cutting themes exhibited low assortativity; Immigration recorded a non-significant coefficient of r = 0.29 (p = 0.095), and Online coordination yielded r = 0.52 (p = 0.011). Bootstrapped temporal evaluations across three-month sliding windows throughout 2019 confirmed that observed variations remained within statistical confidence intervals, demonstrating that these structural segmentation patterns remained stable across the study period.

These quantitative results demonstrate that far-right communication on Telegram operates as a hybrid organizational system characterized simultaneously by shared overarching grievances and sharp structural segmentation. While core ideological doctrines circulate within tightly insulated forwarding clusters resembling selective echo chambers, non-assortative subjects like immigration serve as discursive bridges connecting disparate factions. The author notes several methodological boundaries, including the heavy representation of United Kingdom Brexit channels, the network sparsity resulting from excluding external sources, and the static 2015 to 2019 timeframe which predates subsequent narrative shifts surrounding pandemic conspiracies and QAnon. Nonetheless, the empirical findings show that topical affinities directly organize the structural topology of alternative social networks, providing an empirical framework for examining how digital communities negotiate collective identity, ideological cohesion, and insular content propagation.

Subject of Research: Topic assortativity and message forwarding structures in far-right Telegram networks

Article Title: Forwarding to the choir: topic assortativity in far-right Telegram networks

Article References: Houri, Y. (2026). Forwarding to the choir: topic assortativity in far-right Telegram networks. Applied Network Science, 11(1), Article 45. https://doi.org/10.1007/s41109-026-00796-1

Image Credits: AI Generated

DOI: 10.1007/s41109-026-00796-1

Keywords: applied network science, topic assortativity, Telegram networks, far-right communication, social media polarization, retrieval-augmented generation, multi-label classification, Shannon entropy, message forwarding, homophily, deplatforming, network topology, Louvain community detection, content diffusion, Pushshift dataset, white nationalism, natural language processing