---
title: 'PoliTok-DE: German Political TikTok Corpus'
url: https://www.emergentmind.com/topics/politok-de
type: topic
---

# PoliTok-DE: German Political TikTok Corpus

Searching arXiv for the specified paper and closely related TikTok political communication datasets.
PoliTok-DE is a large-scale multimodal research corpus of TikTok posts related to the 2024 Saxony state election in Germany, designed to support substantive and methodological work on political communication, intolerance, entertainment, and deletion dynamics on short-form video platforms [2509.15860]. The dataset spans video, audio, images, and text, and pairs these modalities with platform metadata and a deletion snapshot indicating whether posts were later removed or made unavailable. It contains 195,373 posts published between 01.07.2024 and 30.11.2024, bracketing the 01.09.2024 election, and identifies 18,842 posts as deleted or unavailable, corresponding to 17.3% of the corpus [2509.15860].

## 1. Dataset scope and political setting

PoliTok-DE focuses on Germany and specifically on the 2024 Saxony (Sachsen) state election context. Its stated purpose is to capture multimodal political discourse around that election and to enable analyses of intolerance and entertainment in political TikTok content, alongside methodological studies of platform policies and deletion behavior [2509.15860]. The dataset is therefore situated at the intersection of computational social science, political communication research, and platform studies.

The political context is integral to the corpus design. The Alternative für Deutschland (AfD) had a strong presence in Saxony and finished first in the 2024 state election, with over one-third of the vote, which motivated attention to right-wing content in the region’s digital sphere [2509.15860]. Within the collected material, AfD is mentioned in 50.9% of all posts and in 70.2% of deleted posts, while the Greens (“Grüne”) appear in 23.8% of all posts; candidate-name keywords were rare, with each occurring in fewer than 1% of posts and some with zero mentions [2509.15860]. A small fraction, 0.8%, of posts returned by the Research API mentioned none of the query keywords, underscoring quirks in API matching [2509.15860].

The operational definition of “political TikToks” in the main collection is keyword-based: posts were treated as political if their description or hashtags matched a curated query set. In the annotated subset, however, “political” was defined more broadly as content that discusses or references policy, political events, actors, or institutions, including news, opinions, advocacy, campaigns, satire, and political memes [2509.15860]. This distinction is methodologically consequential because it separates collection logic from interpretive coding. A plausible implication is that the corpus supports both retrieval-oriented analyses and broader content-analytic approaches to political expression.

## 2. Collection design and query logic

The collection covers posts published between 01.07.2024 and 30.11.2024. TikTok’s Research API was queried daily using a curated keyword list targeting the Saxony election, parties, and prominent local politicians, with a geographic filter `region=DE` to focus on content posted in Germany [2509.15860]. Keywords were matched case-insensitively via exact string matching in post descriptions, including hashtags. The query set comprised the following terms: `['Sachsen', 'ltw24', 'landtagswahl24', 'CDU', 'Afd', 'dielinke', 'grune', 'gruene', 'grüne', 'SPD', 'FDP', 'BSW', 'wahlkampfsachsen', 'Kretschmer', 'JörgUrban', 'joergurban', 'urbanjörg', 'urbanjoerg', 'SusanneSchaper', 'SchaperSusanne', 'StefanHartmann', 'HartmannStefan', 'KatjaMeier', 'MeierKatja', 'WolframGünter', 'GünterWolfram', 'WolframGuenter', 'GuenterWolfram', 'FranziskaSchubert', 'SchubertFranziska', 'PetraKöpping', 'KöppingPetra', 'petrakoepping', 'KoeppingPetra', 'RobertMalorny', 'MalornyRobert', 'SabineZimmermann', 'ZimmermannSabine']` [2509.15860].

The data collection pipeline is described in four stages. First, the project used the Query Videos endpoint at `https://open.tiktokapis.com/v2/research/video/query/` with `region=DE` and the keyword list. TikTok’s documentation indicates that the Research API mirrors website content with a 48-hour lag; based on observed latency, each calendar day was queried 96 hours later to reduce missingness [2509.15860]. Second, full multimodal post media were scraped from the TikTok website immediately after the Research API pulls for each day, together with additional metadata, enabling cross-checks against the API metadata. Third, a single re-scrape was conducted 10 days after the end of the collection window to determine which posts had become unavailable. Fourth, a sample prioritizing deleted posts was labeled for politics, Saxony reference, intolerance, and entertainment [2509.15860].

The paper reports comparison between metadata obtained via the Research API and via scraping, and it notes known inconsistencies in Research API coverage from prior audits. It does not report deduplication procedures or quantified missingness beyond the cross-source comparison and the 96-hour lag collection strategy [2509.15860]. This suggests that PoliTok-DE is intended not only as a substantive corpus but also as a resource for methodological validation across collection channels.

## 3. Multimodal structure and metadata schema

PoliTok-DE is explicitly multimodal. The modalities consist of video as the primary artifact, audio including music, voiceover, and sound effects, images such as thumbnails or covers, and text comprising descriptions, hashtags, and on-screen text visible in the video [2509.15860]. The corpus therefore enables analyses that cannot be reduced to caption text alone.

For each post, the stored metadata include the post identifier; description or caption text, including hashtags; engagement metrics such as views, likes, comments, and shares as provided by the platform at the time of collection; time references, including publication date within the window and collection timestamps; the `DE` region filter; keyword-match indicators; and a deletion flag recorded at the Stage III check [2509.15860]. Web scraping provided the raw media and overlapping metadata for validation. Publicly, however, only post IDs are distributed; full media and metadata are reconstructed through hydration scripts [2509.15860].

The paper does not specify fixed resolutions, frame rates, or codecs. Instead, content is hydrated from TikTok in its native encodings, and researchers are advised to expect TikTok-standard MP4 video and AAC audio, with thumbnails or images in common web image formats such as JPEG or PNG, while recognizing that specific properties vary by post [2509.15860]. Internally, Stage II scraping stored media files keyed by post IDs. Public release via IDs plus hydration code is presented as a compliance-preserving approach that still supports reproducibility, because hydration re-creates the multimodal bundle for each post ID [2509.15860].

A central feature of the schema is the linkage between multimodal artifacts and platform availability state. By combining video, audio, images, textual metadata, and deletion status, the corpus supports analyses of how meaning is distributed across modalities and how that meaning relates to subsequent unavailability. A plausible implication is that PoliTok-DE is especially suited to studying forms of political signaling and intolerance that are only partially expressed in captions.

## 4. Deleted content and deletion dynamics

Deletion status in PoliTok-DE was determined by a single re-scrape 10 days after the end of the collection window, which identified 18,842 deleted or unavailable posts out of 195,373 total posts, yielding an overall prevalence of 17.3% [2509.15860]. The paper uses “deleted/unavailable” as a broad availability category, encompassing posts that were deleted, set to private, or otherwise inaccessible at the follow-up check [2509.15860].

The timing structure matters for interpretation. Initial collection occurred daily, using a 96-hour lag to accommodate Research API mirroring, but deletion was observed only through a one-time follow-up availability check rather than continuous polling [2509.15860]. Consequently, the dataset provides a deletion snapshot rather than a full event history. The paper explicitly notes a core limitation: unavailability can stem from platform enforcement, user deletion, or privacy changes, so deletion does not equate to policy violation [2509.15860]. It also notes that some posts may have been reinstated after the snapshot or deleted only after that check [2509.15860].

The paper gives formulas for prospective analyses enabled by the dataset. Daily deletion rates can be computed as $d_t/n_t$, where $d_t$ is the count deleted by the follow-up check from posts collected on day $t$, and $n_t$ is the total collected on day $t$ [2509.15860]. The cumulative deletion proportion up to $T$ can be expressed as

$$
C(T) = \frac{\sum_{t=1}^{T} d_t}{\sum_{t=1}^{T} n_t}.
$$

For repeated polling designs not reported in the paper, a Kaplan–Meier estimator is described as appropriate:

$$
S(t) = \prod_{i: t_i \le t} \left(1 - \frac{d_i}{n_i}\right),
$$

with hazard defined as $h(t) = f(t)/S(t)$ or in discrete time via event probabilities conditional on survival [2509.15860]. The paper does not present survival or hazard estimates; these formulas are included to guide future work.

This separation between observed snapshot and prospective event-history modeling is important. PoliTok-DE does not itself establish causal reasons for unavailability, but it furnishes post-level date-stamped collections and a later availability state, enabling structured study of deletion-related platform phenomena under explicit observational constraints.

## 5. Annotated subset and coding framework

To illustrate one possible analysis, the paper reports a case study centered on the co-occurrence of intolerance and entertainment using an annotated subset of posts, with deleted posts prioritized in sampling to increase the base rate of intolerance [2509.15860]. The annotation framework assigns binary yes/no labels for politics, Saxony reference, intolerance, and two forms of entertainment.

The coding definitions are precise. “Politics” covers references to policy, politics, or polity, including news, election content, opinions, advocacy, campaigns, and satire. “Saxony” refers to any reference to the federal state, its politics, places, or figures. “Intolerance” is defined as content communicating denial of equal status and refusal to accept different perspectives, cultures, or ideas, following Rossini 2022, with identification guided across modalities including descriptions and hashtags, on-screen text, visuals, and audio. Entertainment follows Oliver and Raney (2011) and is divided into hedonic entertainment, defined as fun, humor, excitement, or pleasure, and eudaimonic entertainment, defined as inspiring, touching, thought-provoking, or meaningful elements [2509.15860].

Each post in the annotated subset was viewed multimodally, and each received between 1 and 4 independent annotations, with a mean of 2.57; annotators could also add optional comments, and the interface displayed media together with codebook tooltips for consistency [2509.15860]. Inter-annotator agreement is reported using agreement percentages and Krippendorff’s alpha. The paper emphasizes that Krippendorff’s alpha is chance-corrected and cites recommended thresholds of at least 0.80 as desirable and at least 0.66 as tentative. It gives the nominal alpha as

$$
\alpha = 1 - \frac{D_o}{D_e},
$$

where $D_o$ is observed disagreement and $D_e$ is expected disagreement by chance. For completeness, it also includes Cohen’s kappa,

$$
\kappa = \frac{p_o - p_e}{1 - p_e},
$$

although $\kappa$ is not reported in the study [2509.15860].

The annotated label distributions and reliability statistics are as follows.

| Label | Distribution and agreement | Krippendorff’s alpha |
|---|---|---|
| Political | Yes 85.7%, No 14.3%; Agreement 94.4% | 0.81 |
| Saxony | Yes 19.2%, No 80.8%; Agreement 90.8% | 0.74 |
| Intolerance | Yes 20.5%, No 79.5%; Agreement 79.7% | 0.48 |
| Hedonic entertainment | Yes 62.9%, No 37.1%; Agreement 75.8% | 0.55 |
| Eudaimonic entertainment | Yes 37.3%, No 62.7%; Agreement 67.2% | 0.38 |

About 13% of annotated posts conveyed both hedonic and eudaimonic entertainment [2509.15860]. The contrast between the relatively high alpha for “Political” and the lower alpha for “Intolerance” and “Eudaimonic entertainment” indicates that some constructs are more difficult to identify consistently in short-form multimodal content. This suggests that implicit and multimodally distributed signals pose nontrivial measurement challenges even when a detailed codebook is available.

## 6. Analytical affordances, access, and reproducibility

The case study is conceptually centered on co-occurrence, particularly the hypothesis that playful framing may co-occur with exclusionary messages in deleted posts [2509.15860]. The paper does not report specific co-occurrence statistics for intolerance with hedonic or eudaimonic entertainment, nor does it report regression coefficients, confidence intervals, or significance tests. Instead, it outlines analytical specifications appropriate for future work. These include the Jaccard index between intolerance $(I)$ and entertainment $(E)$,

$$
J = \frac{|I \cap E|}{|I \cup E|},
$$

mutual information between binary $I$ and $E$,

$$
I(I;E) = \sum_{i \in \{0,1\}} \sum_{e \in \{0,1\}} p(i,e)\log\left(\frac{p(i,e)}{p(i)p(e)}\right),
$$

and a logistic regression specification for an outcome $Y$ such as deletion status or high engagement,

$$
\operatorname{logit}(P(Y=1)) = \beta_0 + \beta_1 I + \beta_2 E + \beta_3 (I \times E) + \sum_k \gamma_k X_k,
$$

where controls $X_k$ can include engagement at collection, account features, or day-of-week [2509.15860]. These formulations define a methodological repertoire rather than reported findings.

Access is organized around policy-compliant hydration. Post IDs are publicly available on Hugging Face at `https://huggingface.co/datasets/tomasruiz/PoliTok-DE`, and hydration code is provided to re-collect media and metadata from TikTok [2509.15860]. The documented workflow is: obtain TikTok Research API credentials and approval; install the hydration environment and dependencies; query the Research API for metadata via the Query Videos endpoint while respecting the 48-hour mirroring delay, with 96-hour delay treated as best practice; run the scraper to fetch multimodal media and cross-check metadata; and optionally re-check availability later to recreate a deletion snapshot [2509.15860]. Access requires approval for the Research API, and TikTok’s documented mirroring delay implies that content removed within 48 hours may not appear in the API [2509.15860].

Deleted content is not publicly distributed in raw form. It can be requested for bona fide research use under strict conditions intended to protect privacy and comply with platform policies; IDs are public, while media must be hydrated directly from TikTok in compliance with TikTok’s Research API policies, terms of service, and applicable copyright and data protection laws [2509.15860]. The release strategy is therefore simultaneously a reproducibility mechanism and an ethical safeguard.

## 7. Ethical framework, limitations, and research uses

The ethical and compliance model of PoliTok-DE is grounded in minimizing legal and ethical risks by excluding raw media from the public release and distributing post IDs with hydration code instead [2509.15860]. This supports reproducibility while avoiding redistribution of user content. The approach is described as respecting TikTok’s Research API rules and German and EU data protection norms by avoiding direct publication of personal media and by constraining access to sensitive deleted content [2509.15860]. The paper does not report formal IRB approval; rather, it emphasizes safeguards and trade-offs in enabling research responsibly while minimizing potential harm [2509.15860].

The limitations are clearly delimited. First, deletion ambiguity means that unavailable content may reflect platform enforcement, user deletion, or privacy changes rather than a single underlying mechanism [2509.15860]. Second, because keyword matching was conducted only in descriptions and hashtags, mentions expressed solely in audio or visuals are not captured, potentially underestimating prevalence [2509.15860]. Third, the single follow-up availability check cannot capture reinstatements or later deletions. Fourth, Research API constraints remain relevant: prior work documents coverage and metric discrepancies, and the 96-hour lag mitigates mirroring gaps but cannot fully eliminate missingness [2509.15860].

Within these constraints, the paper identifies a broad application space. Substantive uses include intolerance research, including implicit and multimodal forms; political communication and party dynamics on TikTok; entertainment-politics interactions; and deletion-related platform phenomena [2509.15860]. Methodological uses include multimodal analyses that combine video, audio, images, and text; the study of deletion dynamics and policy environments; integrated qualitative-quantitative workflows; and validation across API and scraping sources [2509.15860]. In this sense, PoliTok-DE functions not merely as an election dataset but as an infrastructure for examining how political meaning, platform governance, and multimodal communication interact on short-form video systems.

Source: https://www.emergentmind.com/topics/politok-de