Conference Agenda

The study of phenomena related to public opinion online and especially political polarization garners significant interest in Computational social sciences. The undertaking of several studies of political phenomena in social media mandates the operationalization of the notion of political stance of users and contents involved. Relevant examples include the study of segregation and polarization online, or the study of political diversity in content diets in social media. While many research designs rely on operationalizations best suited for the US setting, few allow addressing more general design, in which users and content might take stances on multiple ideology and issue dimensions, going beyond traditional Liberal-Conservative or Left-Right scales. To advance the study of more general online ecosystems, we present a methodology for the computation of multidimensional political positions of social media users and web domains. We perform a case study on a large-scale X/Twitter population of users in the French political Twittersphere and web domains, embedded in a political space spanned by dimensions measuring attitudes towards immigration, the EU, liberal values, elites and institutions, nationalism and the environment. We provide several benchmarks validating the positions of these entities (based on both LLM and human annotations), as well as a discussion of the case studies in which they can be used, including, e.g., AI explainability, political polarization and segregation, and media diets. To encourage reproducibility and further studies on the topic, we publicly release our anonymized data.

What is the structure of language interactions? We already know that language is heavily structured by Zifp’s law (Zipf 1935), be it at the general vocabulary level (Condon 1928), for higher n-grams (Ha et al. 2009), or over the syntactic dependency network (Ferrer i Cancho et al. 2004). This behavior is also found at a more local level: for lexical niches like natural entities or numbers (Piantadosi 2014), for individual semi-schematic constructions (Feltgen 2020), for argument structure constructions (Ellis 2012), etc. The emergence of Zipf’s law at the general level is believed to reflect a sharing of the coding and decoding efforts between speaker and hearer (Ferrer i Cancho & Solé 2003), while Zipf’s law at the individual level has been related to learning mechanisms (Goldberg et al. 2004) and to the scale-free structure of semantic networks (Ellis et al. 2014). However, these works do not address the structure of the interaction between two such Zipfian paradigms, like the noun and adjective ones in the epithet construction, even though combination is crucial to generate a meaningful, creative, and diverse language output.

In this contribution, we aim to address general properties of the syntactic interactions between two paradigmatic slots. To do so, we shall focus on the epithet construction in contemporary French (1980-2024), based on data extracted from the Frantext corpus associated with that period (ATILF 1998-2025). To reduce the volume of occurrences and ensure a better homogeneity in the output, we focus on indefinite contexts (e.g. une joie sincère), yielding 171,000 interactions between 12,000 nouns and 9,000 adjectives, for a total of 111,000 different combinations.

Most empirical studies on online public opinion rely on a set of preselected keywords related to a specific topic. In contrast, this study adopts a fully automated approach based on Twitter (now X) data collected between September 2021 and June 2022. We construct a weighted hashtag network, where the weight of a link between two hashtags represents the number of unique users who used both hashtags in the same tweet. Discussion topics are identified by detecting communities within this semantic network.

This method has previously been used to analyze political discourse during elections in Argentina and to study interactions between The New York Times and its Twitter followers. However, those studies used static networks built from the full data set, incorporating information from the future—making them unsuitable for real-time event tracking.

This work introduces a method to overcome that limitation by analyzing evolving semantic networks. Two network-building procedures are compared. The first is cumulative: it starts with data from the first month and adds new data weekly, preserving memory of all past discussions. The second uses a sliding window of one week, where new data replaces the oldest, causing a weekly loss of memory.

For each weekly network, discussion topics are extracted via community detection. Then, for any group (e.g., media followers or political supporters), a vector is constructed representing their participation level in each topic. The similarity between groups is measured by comparing their vectors. The study finds that both procedures generally yield similar trends in the evolution of group similarities, with a few exceptions in unusual cases.

Moreover, because the semantic landscape is updated weekly, the method also tracks how topics themselves evolve over time. Using a dynamic community tracking approach, the researchers observe how topics emerge, fade, split, or merge. For instance, in controversial debates like vaccination policies, certain topics gradually absorb new and highly used hashtags that lean toward more extreme positions.

Models of opinion dynamics describe how opinions are shaped in various environments. While these models are able to replicate macroscopical opinion distributions observed in real-world scenarios, their capacity to align with data at the microscopical level remains mostly untested. We evaluate the capacity of the celebrated voter model to capture individual opinions in a fine-grained Twitter dataset collected during the 2017 French Presidential elections. Our findings reveal a strong correspondence between individual opinion distributions in the equilibrium state of the model and ground-truth political leanings of the users. Additionally, we demonstrate that discord probabilities accurately identify pairs of like-minded users. These results emphasize the validity of the voter model in complex settings, and advocate for further empirical evaluations of opinion dynamics models at the microscopical level.

Telegram has evolved beyond a simple messaging app to become a major online social network. It is structured around two types of channels: broadcast channels, where only admins can post, and discussion channels, where all members can participate. While its unique structure is noteworthy, Telegram has primarily attracted academic attention due to its role in spreading hate speech and disinformation. Despite this, the platform remains under-studied compared to other social media.

This study aims to deepen understanding of how Telegram's network grows and how information spreads on it. Using the Pushshift Telegram dataset, which includes 29,000 public channels, 2 million users, and around 300 million messages (from September 2015 to October 2019), the authors created a temporal network of over 7.5 million edges. Each edge represents a message being forwarded from one channel to another, capturing the flow of information.

We first aggregated this data into a static, weighted network for topological analysis. This revealed classic social network features: scale-free distributions (indicating a few highly connected hubs), high clustering, and assortativity (channels with similar attributes, such as language, tend to connect). Community detection further confirmed these patterns.

We also examined the temporal dynamics of forwarding. The time between forwarding events follows a piecewise power law, with different behavior for intervals shorter or longer than a day. This pattern held across activity levels, suggesting a bursty communication style typical in human interactions. The burst train size distribution further confirmed this burstiness.

Building on these findings, we developed a growth model for the Telegram network. It incorporates key dynamics: bursty message forwarding driven by memory effects, and network growth shaped by focal and triadic closure (i.e., users forming connections based on shared interests or mutual connections). We adapted existing topological and temporal models to accurately reflect both the structure and temporality observed in the data.

Ultimately, this work offers valuable insights into how information spreads on Telegram and lays the foundation for future studies—particularly those exploring the effects of external interventions on these dynamics.