Quantifying Political Partisanship for Cross-Platform Analyses

Authors: Mia Ameen, Christopher G. Healey Affiliation: North Carolina State University Venue: arXiv cs.SI, 2026 — arXiv:2607.21842

TL;DR

Political partisanship measurement in social media typically relies on platform-specific features (follower networks, community structure), making cross-platform comparison difficult. This work proposes a text-based partisanship measurement method grounded in external media bias ratings, and validates it on ~1.3 million posts from Bluesky and Truth Social during the 2024 U.S. presidential election, demonstrating that the approach captures platform-specific partisan dynamics comparably across different social networks.

Contributions

• A text-based, demographically-informed partisanship measurement method using semantic embeddings and external media bias scores of cited news sources.
• First cross-platform comparative analysis of political discourse on Bluesky and Truth Social, showing platform-specific partisan patterns during the 2024 election cycle.
• Validation framework demonstrating that the constructed partisanship axis generalizes beyond the seed platforms and captures within-platform partisan distinctions comparably.

Method

Partisanship measurement pipeline:

1. Text embedding: All posts are encoded using OpenAI's text-embedding-3-small model with semantic deduplication to reduce redundancy.
2. News source identification: Posts citing news sources are identified; AllSides Media Bias ratings provide ground-truth partisan leanings for these sources (ranging from far-left to far-right).
3. Clustering: Posts and their cited news sources are clustered using semantic similarity, with cluster-level aggregation producing separate left- and right-leaning clusters.
4. Axis construction: Following Waller and Anderson (2021), a partisanship dimension is constructed as the vector difference between the centroids of left- and right-leaning clusters, normalized via L2 normalization.
5. Post scoring: Each post's embedding is projected onto this axis; the resulting score reflects alignment with left- or right-leaning partisan positions.

Validation:

• In-distribution validation: Posts citing news sources with known AllSides bias scores are tested for correlation between predicted partisanship score and the source's published bias rating.
• Out-of-distribution validation: Posts without cited news sources (drawn from a separate 2024 U.S. election dataset) are scored to test generalization beyond the training platform.
• Seed pair sensitivity analysis: The choice of initial left/right-leaning seed posts is tested via 630 possible cluster pairings; the method is robust across seed selection.
• Qualitative analysis: Manual inspection of highest- and lowest-scoring posts confirms the axis captures meaningful partisan themes (Democratic criticism, anti-Democratic rhetoric, pro-Palestinian stances, immigration framing, etc.).

Data

• Truth Social: 1.2 million posts collected Feb–Nov 2024 using the Truth Social API; dataset from Shah et al. (2024).
• Bluesky: Approximately 644,000 posts collected May–Oct 2024 via the AT Protocol Python SDK.
• Temporal scope: Six months preceding the 2024 U.S. presidential election, peak political discourse period.
• Preprocessing: Removal of posts with missing/empty content, deduplication, filtering for substantive content (≥3 tokens), HTML/emoji cleanup, removal of reposts (retweets on Truth Social).

Results

• Within-platform partisan distribution: Both Truth Social and Bluesky exhibit internally coherent left-leaning and right-leaning clusters with significant partisan separation.
• Cross-platform partisan asymmetry: Truth Social shows greater prevalence of right-leaning discourse (median partisanship score closer to positive, supporting earlier findings on platform ownership and moderation differences). Bluesky exhibits broader ideological range but with notable left-leaning tendency in the aggregate.
• Platform-specific linguistic norms: The partisanship axis captures substantively different discourse patterns on each platform; for example, immigration rhetoric differs in framing and emphasis between Truth Social and Bluesky.
• News source citations: Distinct news source preference patterns across platforms—Truth Social users cite right-leaning outlets (RawStory, New York Times, The Guardian most cited), while Bluesky users show broader mainstream and progressive outlet citation.
• Robustness: Sensitivity analysis confirms the axis construction is not unduly dependent on seed pair selection; correlations with media bias scores remain stable across alternative seed configurations.

Connections

• Related to Mosleh et al. on partisan homophily via shared focus on partisanship as a structural force in social networks.
• Methodologically related to Waller & Anderson's approach to embedding-based social dimension construction.
• Complements Political polarization and ideological echo chambers and Social media polarization and echo chambers topics; provides quantitative measurement approach for understanding partisan sorting across platforms.
• Relevant to Echo Chambers Polarization and Filter Bubbles as it documents platform-specific partisan clustering during high-salience election period.

Notes

Strengths: The method sidesteps platform-specific structural features (follower networks, community APIs) that limit cross-platform comparison. Grounding in external media bias scores provides an interpretable, externally-validated anchor for the partisanship axis. The application to two ideologically distinct platforms (Truth Social as right-leaning, Bluesky as alternative/progressive) during a high-salience political event (2024 election) provides a natural quasi-experimental contrast.

Limitations and questions: The method depends on external media bias ratings (AllSides), which themselves may be contested or biased. Posts without news citations are scored post-hoc without the same level of validation. The temporal window (May–Oct 2024 for Bluesky, Feb–Nov for Truth Social) is election-specific; generalization to non-election periods or other political contexts remains unclear. The choice of semantic embedding model (OpenAI) and clustering hyperparameters (k-means elbow, similarity threshold) may affect replicability; no ablation studies are provided on these choices.

Relevance to misinformation: While not a disinformation-detection paper per se, partisanship measurement is foundational to understanding how partisan identity shapes information acceptance and resistance—key mechanisms in misinformation research. The cross-platform approach is increasingly important as researchers study how alternative platforms (Truth Social, Mastodon, etc.) differ from mainstream outlets in their information ecosystems.