ACL paper: Russia disinfo tailors framing by target language
TL;DR
- Sinelnik and Hovy analyze roughly 8,000 news articles from eight years of Russia-backed disinformation campaigns spanning four languages and fifteen target countries.
- The paper finds campaigns consistently favor specific frames depending on the audience's language, and Russian-language coverage varies frames by target region.
- The two most prominent automatic frame-analysis models underperform and show high disagreement, exposing a gap in multilingual detection tooling.
A quietly useful ACL 2024 Student Research Workshop paper from Antonina Sinelnik and Dirk Hovy, published on the ACL Anthology, takes computational framing analysis out of its usual English-only comfort zone and points it at eight years of Russia-backed disinformation. The corpus is around 8,000 news articles across four languages, targeting fifteen countries. That scale is modest by web-scraping standards but unusual for framing work, which tends to live on smaller English datasets.
The headline finding is that the campaigns are not just translated. According to the paper, disinformation "consistently and intentionally" favors specific frames depending on the target language of the audience, and Russian-language articles further shift which frames they emphasize based on the region being covered. In other words, the same underlying story gets a different rhetorical shape depending on who is meant to read it, which matches what analysts of state information operations have long suspected but is worth having in a measured, peer-reviewed form.
The less flattering finding, and arguably the more actionable one for anyone building trust-and-safety tooling, is about the tools themselves. The authors report that the two most prominent models for automatic frame analysis underperform on this multilingual material and show high disagreement with each other. If you were planning to bolt an off-the-shelf frame classifier into a monitoring pipeline for non-English content, that is a warning shot.
The honest caveat is that this is a short student-workshop paper, and the summary I retrieved does not name the specific four languages, the fifteen target countries, or which two frame-analysis models were evaluated and by which metric. Those are exactly the details a practitioner would want before generalizing the result, so treat the specifics as reported, not settled.
The forward-looking part is straightforward: there is now a published, empirical gap between what multilingual disinformation is doing and what current framing tools can see. That is a good prompt for the next round of benchmarks, and a good reason for newsrooms and OSINT teams tracking Russia-linked operations to assume their non-English monitoring is noisier than their English monitoring.
Shared on Bluesky by 2 AI experts
-
#TBT #NLProc "Narratives at Conflict" by Sinelnik and @dirkhovy.bsky.social looks at hidden tactics of disinformation campaigns! They analyzed 8,000 news articles across 4 languages to reveal how disinformation campaign…
View on Bluesky →
Originally reported by aclanthology.org
Read the original article →Original headline: Narratives at Conflict: Computational Analysis of News Framing in Multilingual Disinformation Campaigns