Who leads? Who follows? Temporal dynamics of political dogwhistles in Swedish online communities

Max Boholm, Gregor Rettenegger, Ellen Breitholtz, Robin Cooper, Elina Lindgren, Björn Rönnerstrand, Asad SayeedACL 2025

TL;DR

This paper investigates how political dogwhistles spread across online communities, finding that more radical communities (like Flashback) often precede less radical ones (like Familjeliv) in adopting these coded messages. By analyzing time-series data, they demonstrate a directional influence, suggesting a "mainstreaming" process where dogwhistles move from fringe to more mainstream discussions.

What Problem Does This Paper Address?

Political dogwhistles are a significant challenge for democratic discourse. They allow politicians and groups to convey controversial or exclusionary messages to a select in-group while appearing to communicate something innocuous to the general public. This ambiguity makes it difficult to hold them accountable and can subtly normalize harmful ideologies. While dogwhistles have been studied, their temporal dynamics – how they emerge, spread, and change over time, especially across different online communities – remain underexplored. This paper aims to fill this gap by examining the temporal relationship between dogwhistle usage in a more radical online community and a less radical one, specifically in the Swedish context.

Methods

The researchers employed a multi-faceted approach:

  • Lexical Semantic Change (LSC) Modeling: They used Sentence-BERT (SBERT) to create sentence embeddings. The "intensity" of a dogwhistle expression (DWE) at a given time was defined as the similarity between its contextual embedding and the embedding of its in-group meaning. This allows for quantifying how strongly a DWE is being used with its coded, in-group meaning.
  • Vector Autoregression (VAR) Models: To understand the temporal dynamics and directional influence between dogwhistle intensity in two communities, they used VAR models. These models analyze how past values of a time series (dogwhistle intensity in one community) predict future values of another time series (dogwhistle intensity in another community).
  • Granger Causality Tests: To more rigorously test for temporal precedence, they applied Granger causality tests. This statistical method determines if past values of one time series can significantly improve predictions of another time series.
  • Data Imputation: To handle periods where specific dogwhistles were not observed in a community (zero frequency), they used two imputation methods: mean imputation and Last Observation Carried Forward (LOCF). This was crucial for building continuous time series for analysis.

Datasets

The study utilized two primary datasets:

  • Swedish Online Discussion Forums:
    • Flashback: A large, well-known Swedish forum with a strong anti-immigrant subtext, known for hosting controversial discussions.
  • Familjeliv: A Swedish forum focused on family and everyday life, generally less political and contentious than Flashback.
  • Corpus data from these forums was collected quarterly from 2010 to 2024.
  • Word Replacement Survey Data: Data from Lindgren et al. (2024) was used to validate the identification of three Swedish dogwhistle expressions (DWEs) and to model their in-group meanings. This survey involved Swedish citizens replacing potential DWEs in sentences with words that preserved the sentence's meaning.

Key Insights and Ideas

  • Directional Influence: The most striking finding is the evidence for a directional flow of dogwhistles. Intensity changes in the more radical Flashback community consistently predict changes in the less radical Familjeliv community, but not vice-versa. This suggests that radical communities act as a "source" or "gateway" for dogwhistles.
  • Mainstreaming of Far-Right Discourse: The findings strongly support the idea that dogwhistles are a mechanism for the mainstreaming of far-right or exclusionary ideologies. They move from fringe spaces into broader public discourse.
  • Temporal Dynamics Matter: Analyzing dogwhistles not just as static phenomena but as dynamic processes over time reveals crucial insights into their diffusion and impact. The "fluctuating pattern" of intensity, rather than a simple trend, highlights the influence of external events and internal community dynamics.
  • Community Specificity: The study underscores that dogwhistle evolution is community-dependent, with different rates of change and patterns observed across different online spaces.

Key Results

  • Hypothesis H1 Supported: Changes in dogwhistle intensity in Flashback significantly predict changes in Familjeliv. This was particularly evident for the DWEs "berika" (enrich) and "återvandring" (re-migration).
  • Hypothesis H2 Supported: Changes in dogwhistle intensity in Familjeliv did not significantly predict changes in Flashback.
  • Partial Support for "Globalist": While the general pattern held for "globalist," the relationship was less consistent, potentially due to its strong pre-existing association with anti-Semitic conspiracy theories.
  • Granger Causality Confirmed: Granger causality tests confirmed that Flashback significantly Granger-causes Familjeliv for "berika" and "återvandring," reinforcing the directional influence.
  • Autocorrelation: Both communities exhibited strong autocorrelation, meaning past changes in intensity within a community were good predictors of its own future intensity.

Limitations and Future Work

  • Data Scarcity: The authors acknowledge that only a few dogwhistles are "active" at any given time, limiting the number of DWEs that can be studied longitudinally. They note that finding more DWEs requires extensive selection processes.
  • Zero-Frequency Quarters: The imputation of data for quarters where DWEs were absent in Familjeliv (especially for "globalist" and "återvandring") is a concern for robustness. While "berika" had minimal imputation, the results for the other DWEs should be interpreted with caution.
  • Interpreting Absence: The paper notes that the absence of a term (zero frequency) is itself an observation, reflecting lower probability of DWE usage in communities less sympathetic to the in-group meaning.
  • Future Work: The authors suggest exploring what drives the intensity patterns in influential communities like Flashback, positing that prominent events (elections, refugee flows) and the salience of political issues play a role. They also call for further research into the "leader-follower" relationship and the mechanisms of diffusion, including how users migrate between communities and platforms. Investigating the interaction of influential users, extremist infiltration, and agenda-setting by media could also be fruitful.

Summary updated

BibTeX
@inproceedings{boholm2025who,
  author = {Max Boholm and Gregor Rettenegger and Ellen Breitholtz and Robin Cooper and Elina Lindgren and Björn Rönnerstrand and Asad Sayeed},
  title = {{Who leads? Who follows? Temporal dynamics of political dogwhistles in Swedish online communities}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2025},
}
Figure