💬 NLP

Conflict or Strategy? Asymmetric Role Framing of La France insoumise and Rassemblement National in French News Headlines, 2022-2025

This study analyzes 28,592 French news headlines from 2022 to 2025 to reveal that media framing of left-wing La France insoumise and right-wing Rassemblement National is characterized by a stable role asymmetry—portraying LFI primarily through a conflict register and RN through a strategic-electoral register—rather than by symmetric delegitimization, while also establishing a construct-stratified reliability framework for LLM-based political text annotation.

Original authors: Amr Sobhy

Published 2026-08-12

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Amr Sobhy

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are walking into a giant, noisy cafeteria where everyone is shouting about who is winning the game. In the world of science, there is a whole team of researchers dedicated to listening to these shouts to figure out what the "news" is actually saying. They study something called framing. Think of framing like the camera lens a photographer uses. If you take a picture of a messy room through a "chaos" lens, it looks like a disaster; through a "creative process" lens, it looks like an artist at work. The same room, two different stories. Another key idea is agenda-setting, which is simply the art of deciding which topics get the most airtime. But this paper asks a deeper question: When the news talks about two rival political teams, does it treat them like mirror images of each other (two "extremes" causing the same kind of trouble), or does it cast them in totally different roles, like a "villain" versus a "strategist"? This matters because how we are told to see our leaders changes how we vote, how we argue, and how we understand our own democracy.

This paper, written by Amr Sobhy from the French News Lab, dives into 28,592 French news headlines from 2022 to 2025 to see how the press describes two major challenger parties: La France insoumise (LFI), a radical-left group, and the Rassemblement National (RN), a nationalist right-wing group. The researchers didn't just read these headlines; they built a high-tech "robot jury" using three different powerful AI models to label every single headline. They asked the robots to decide: Is this headline about a fight (conflict)? Is it about a chess match or election strategy (strategic game)? Is one party being painted as the bad guy (delegitimization)?

The biggest surprise? The news doesn't treat these two parties as twins. Instead, it gives them very different jobs in the story. LFI is almost always cast as the fighter or the aggressor. When LFI is in the news, the headlines are more likely to say they are "attacking," "clashing," or "denouncing" something. It's like the news is constantly showing LFI throwing punches. On the other hand, RN is cast as the strategist. Headlines about RN are much more likely to talk about "elections," "alliances," and "tactics." The news treats RN like a team playing a calculated game of chess, while LFI is the team screaming in the ring.

The study found that this "Fighter vs. Strategist" split is very strong and consistent. About 13.8% of LFI headlines used a conflict frame, compared to only 8.7% for RN. Conversely, 45.9% of RN headlines used a strategic-game frame, while only 37.2% of LFI headlines did. The researcher is very confident about this specific finding because the three different AI models agreed with each other, and a team of human experts double-checked the work and confirmed the pattern.

However, the paper is careful to say what it doesn't know. It explicitly rules out the idea that the news is simply "negative" toward both parties equally. The negativity isn't the main story; the type of story is. The paper also suggests, but cannot prove, that this might be because LFI actually does more fighting (they led major protests against pension reforms), while RN is actually doing more strategizing (trying to normalize themselves for elections). The study is like a camera recording a scene; it can tell us exactly what the camera saw, but it can't tell us if the actors were improvising or following a script.

There is one tricky part: the "moral" judgment. The researchers tried to see if the news was calling one party "illegitimate" (a fake player) or "blameworthy" (the one at fault). Here, the results were messy. The AI models disagreed a lot, and the pattern changed depending on which newspaper you looked at. Left-leaning papers tended to paint RN as the villain, while right-leaning papers did the opposite. Because the data was so split, the author says we can't make a single, solid claim about who the whole French press thinks is "bad." They call this a "structured tendency" rather than a hard fact.

In short, the paper reveals that French headlines don't just say "these two parties are bad." They say, "One party is the chaotic brawler, and the other is the calculating player." This distinction is stable across time and across most newspapers, suggesting that the French press has settled into a specific way of telling the story of these two rivals, casting them in roles that are as different as a boxer and a chess master.

Technical Summary: Conflict or Strategy? Asymmetric Role Framing of La France insoumise and Rassemblement National in French News Headlines, 2022–2025

Problem and Research Gap
This study addresses a limitation in computational framing research: the tendency to conflate "selection" (coverage volume) with "salience" (how issues are framed) and the reliance on generic valence measures (sentiment, negativity) that fail to capture the specific political roles assigned to actors. While existing literature (e.g., Entman, 1993; Semetko and Valkenburg, 2000) establishes that conflict and strategic-game frames are common, no prior computational design has systematically tested whether left- and right-populist challengers are framed as symmetric "extremes" or as fundamentally different political adversaries within the same media system. Specifically, the paper investigates whether French headlines assign distinct behavioral and normative roles to La France insoumise (LFI) and the Rassemblement National (RN) between 2022 and 2025.

Methodology
The research utilizes a large-scale observational design analyzing 28,592 French news headlines from 25 outlets (ranging from far-left to far-right) published between January 2022 and December 2025. The corpus was filtered to include headlines with surface references to LFI, RN, their coalitions, or key leaders.

Annotation Pipeline: The study employs a three-model majority-vote pipeline using open-weight Large Language Models (LLMs): GPT-OSS-120B, Llama-3.3-70B-Instruct, and Mistral Large 2. These models were queried with identical prompts containing strict boundary conditions to distinguish behavioral constructs (e.g., conflict requires an explicit confrontation verb) from normative judgments.
Validation: To address the risk of correlated bias and circularity (where models might simply replicate training data patterns), the pipeline was validated against a stratified human audit of 400 headlines. The audit specifically oversampled disagreement-heavy cases.
Statistical Analysis: Framing rates were analyzed using logistic regression with outlet and year fixed effects. Robustness was verified through cluster-bootstrapped confidence intervals (accounting for outlet clustering) and 10,000-shuffle permutation tests.
Construct Hierarchy: The author establishes a two-tier reliability framework:
1. Behavioral Tier: High agreement (79.6–85.5% unanimous model agreement; $\kappa \ge 0.76$ human validation). Includes Conflict Frame and Strategic-Game Frame.
2. Moral-Accounting Tier: Lower agreement (53.4% unanimous; $\kappa = 0.644$ ). Includes Legitimacy (delegitimization/normalization) and Blame. These are treated as "structured tendencies" rather than stable properties.

Key Results
The study finds a robust role asymmetry rather than a simple valence asymmetry (i.e., one party being more "negative" than the other).

Core Behavioral Asymmetry:
- LFI is framed significantly more often through a conflict register (13.8% of headlines vs. 8.7% for RN; Adjusted OR = 0.614). This persists across years and outlet families.
- RN is framed significantly more often through a strategic-electoral register (45.9% vs. 37.2% for LFI; Adjusted OR = 1.414). This advantage is strongest in 2023 but diminishes in 2025 following legal events affecting RN leadership.
- Aggressor Role: LFI is more frequently cast in the syntactic role of the Aggressor (initiating confrontation), though this is treated as corroborating syntax due to lower validation reliability compared to the binary behavioral frames.
Moral Accounting and Outlet Structure:
- Delegitimization: While pooled data suggests RN is more frequently delegitimized (31.2% vs. 25.7%), this finding is highly outlet-structured and temporally unstable. Left-leaning outlets (e.g., Libération) strongly delegitimize RN, while right-leaning outlets (e.g., Le Figaro, JDD) normalize it or even delegitimize LFI. The aggregate gap is an "aggregation null" that masks polarized editorial logics.
- Blame and Target: Similar to delegitimization, blame attribution shows strong outlet-level divergence (left outlets blame RN; right outlets blame LFI), resulting in near-zero aggregate effects.
- Stability: The behavioral asymmetry (LFI = Conflict, RN = Strategy) is direction-stable across all three annotation models and survives bootstrap and permutation tests. The moral-accounting layer is sensitive to model disagreement and outlet ideology.
Temporal Dynamics:
- The strategic framing of RN peaked in 2023 (non-election year) and collapsed in 2025 following Marine Le Pen's conviction, which shifted coverage toward judicial registers.
- LFI's conflict framing remained consistent, peaking in 2025 during the aftermath of the pension reform protests (Article 49.3).

Contributions and Significance
The paper makes three primary contributions to computational social science and political communication:

Political-Role Assignment as a Construct: It operationalizes "political-role assignment" (e.g., aggressor, strategic competitor) as a distinct analytical target, moving beyond generic sentiment or negativity. This allows for the decomposition of frames that valence-based measures conflate.
Construct-Stratified Reliability Framework: The study establishes a methodology for calibrating majority-vote LLM pipelines by distinguishing between high-stability behavioral constructs and lower-stability normative constructs. It demonstrates that while LLMs can reliably detect behavioral registers (conflict/strategy), they struggle with contested normative judgments (legitimacy/blame) without explicit human validation.
Empirical Evidence of Asymmetric Framing: It provides evidence that the French press does not treat LFI and RN as symmetric "extremes." Instead, it assigns them divergent political roles: LFI is narrated as a behavioral challenger engaged in conflict, while RN is narrated as a strategic electoral competitor. This differentiation persists even when controlling for outlet ideology, suggesting a structural editorial differentiation rather than simple partisan bias.

Limitations and Cautions
The author explicitly notes that the observational design cannot distinguish between "behavioral supply" (parties actually generating different types of events) and "framing construction" (journalists applying different templates). The study documents the distribution of roles in headlines, not the independent construction of those roles by the press. Furthermore, the reliance on headlines limits the scope to high-salience framing units, and the use of LLMs introduces a risk of circularity if models have internalized the very press patterns they are measuring, though the stability of results across models and human audits suggests the signal is robust.

More like this