Multi-platform analysis of electoral discourse on social media as a research infrastructure problem
Summary
Using the 2025 German federal election as a case study, this paper recasts multi-platform electoral discourse analysis as a research infrastructure problem rather than a sequence of standalone empirical studies. The authors collected 81,866 X posts and 43,597 TikTok videos across seven German parties, applied BERTopic for topic modeling, and built a multimodal pipeline combining facial, vocal, and textual analysis on diarized TikTok segments. They argue that the central obstacle to studying electoral harms is not a lack of studies but the absence of consolidated, reusable tools, workflows, and documentation for cross-platform observability under increasingly constrained platform access. The paper closes with four recommendations on data standardization, quality, content-type-tailored analysis, and interdisciplinary collaboration.
Key Contributions
- Reframes multi-platform election analysis as an infrastructure problem requiring reusable foundations rather than ad hoc empirical projects.
- A documented parallel X/TikTok collection case study, distinguishing “authored” (official accounts) from “promoted” (mentions, hashtags, retweets) content.
- A concrete multimodal TikTok pipeline integrating facial emotion (DeepFace), vocal tone (HuBERT), and textual sentiment (XLM-R) at the diarized-segment level.
- A catalogue of biases — algorithmic, temporal, sampling, and noise — arising from post-API scraping conditions.
- Four actionable recommendations: standardized data structures, higher data-quality benchmarks (e.g., transparent electoral account labeling), content-type-tailored analysis, and interdisciplinary expertise integration.
Methods
The study covers 6 Nov 2024 (collapse of the German government) to 23 Feb 2025 (election day). X data was gathered via an API v2-based scraper seeded with official party accounts and expanded through mentions; TikTok data was assembled by aggregating partial results across repeated hashtag-based calls and expanded via party-linked accounts. Posts were clustered with BERTopic and tracked temporally. TikTok videos were diarized with pyannote.audio, then analyzed for facial emotion (OpenCV + DeepFace), vocal tone (HuBERT), and textual sentiment (XLM-R), with Cramér’s V quantifying inter-modal associations.
Findings
- Dataset: 81,866 X posts from 541 accounts and 43,597 TikTok videos from 4,682 accounts across seven parties.
- Parties strategically diverge across platforms: BSW dominates on X but barely uses TikTok; AfD is highly active on TikTok; FDP and CDU span both.
- Facial, vocal, and textual modalities on TikTok are only weakly associated (Cramér’s V ≤ 0.104), functioning as semi-independent channels.
- Political TikTok content frequently pairs negative/critical text and visuals with calm, neutral vocal delivery — a politainment signature.
- Russia/Ukraine and immigration top both platforms but in reversed order; migration is more salient on TikTok, while Israel/Gaza and EU topics dominate on X.
- About two-thirds of hashtag-scraped TikTok content was non-political, signaling substantial noise bias.
- X’s ~1,000 posts/day per-account API ceiling prevents historical recovery for active users; TikTok returns algorithmically filtered partials, embedding temporal and engagement bias.
Connections
This paper sits squarely in the “post-API age” debate over what computational election research can still do, connecting to Freelon2024-sc, Bruns2025-fz, and Bruns2026-yv on access decay, and to Rieder2025-ju and Rieder2026-pp on platform observability as infrastructure. Its multimodal TikTok pipeline and emphasis on TikTok-specific scraping artefacts connect to Achmann-Denkler2026-lx, Pierri2025-hm, and Bouchaud2026-lr, while its cross-platform German election framing relates to Heiss2026-qv and the DSA Article 40 vetted-access discussions echoed in Ohme2026-nv and Vincent_undated-re.
Podcast
A research-radio episode discusses this paper: Listen