The 2017 French presidential election was a landmark event, characterized by a shift from the traditional left-right binary to a complex multi-party landscape. To capture this complexity, the #Élysée2017fr dataset provides a massive, expert-annotated corpus of over 22,000 Twitter profiles and nearly 10 million interactions. This dataset moves beyond automated labeling to offer a high-quality "ground truth" for studying political stance, echo chambers, and misinformation.
Addressing the Limitations of Political Datasets
Most research into political stance detection on social media relies on datasets that are either too narrow or too noisy. Many existing datasets focus on binary oppositions (like Democrats vs. Republicans in the US) or rely on automated labeling—such as assuming a retweet equals support—which fails to account for sarcasm, criticism, or undecided voters.
The 2017 French election featured five major parties ranging from far-left to far-right, making it an ideal case study for multi-class stance detection. The authors identified a need for a manually verified dataset that reflects this real-world complexity to help researchers build better models for echo chamber analysis and the study of digital political discourse.
Methodology: Expert-Driven Annotation
The authors used a multi-step pipeline to ensure data quality:
- Data Collection: Monitoring keywords related to candidates and parties from November 2016 to May 2017, capturing over 42 million tweets.
- Filtering: Retaining active users (at least 10 posts) who mentioned a political party in their bio, handle, or name.
- Expert Annotation: Instead of crowdsourcing, 16 domain experts familiar with the French political landscape manually verified profiles.
- Multi-Label Approach: Annotators allowed for "overlapping" affiliations. If a user expressed support for two parties, they were tagged with both, capturing the nuances of "swing voters."
- Validation: Cross-annotation of 1,000 profiles yielded a high consensus score ($A_m = 0.75$), validating the reliability of the process.
Dataset Composition and Engagement
The primary output is the #Élysée2017fr dataset, which includes 22,853 annotated profiles, 2.4 million original tweets, and 7.7 million retweets. A key finding is the disproportionate activity levels across different political affiliations.
As shown in Figure 1, while La France Insoumise (FI) and Les Républicains (LR) had the highest profile counts (22% and 19% respectively), the Front National (FN) exhibited the most "aggressive" amplification behavior. Despite representing only 15% of the users, FN profiles generated 20% of the total retweets. In contrast, the centrist En Marche! (EM) and traditional Parti Socialiste (PS) showed lower relative engagement metrics.
Temporal Dynamics of the Campaign
Twitter activity was closely tied to real-world political events. The dataset tracks these fluctuations, providing a timeline of digital engagement that mirrors the campaign's intensity.
Figure 3 highlights sharp peaks during televised debates and the two rounds of voting. Specific events, such as "FillonGate" or the publication of Macron's program, triggered distinct surges in activity for the respective parties. Notably, the FN showed a unique concentration of activity, with 26% of their total tweet volume occurring in the final week of the campaign.
Key Insights: Network Segregation and International Interest
Analysis of the interaction networks revealed a "porous" mention network. While the retweet network is highly segregated (users mostly amplify their own side), the mention network is much more integrated. This suggests that while users reside in ideological echo chambers for content amplification, they frequently engage with opponents through mentions, often for debate or criticism.
Furthermore, the dataset highlights international interest in the election. The far-right FN had the highest percentage of profiles located outside of France (11%), nearly double that of other parties, suggesting a broader international network of support or interest.
Conclusion and Future Work
The #Élysée2017fr dataset offers a rich resource for researchers. While it reflects the demographic biases of Twitter (younger, more male), its expert-verified labels provide a robust foundation for training N-class stance detection models and studying the diffusion of misinformation. Future work can leverage this "ground truth" to identify influential nodes and analyze ideological shifts among the 803 profiles identified with overlapping affiliations.
Summary updated