Pith. sign in

REVIEW 4 major objections 6 minor 48 references

This paper claims that platform-level incentives—not just topic choice—shape climate discourse online: paid Meta ads concentrate on promoting specific solutions in a formal, forward-looking register, while Bluesky posts concentrate on syste

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Paid climate ads on Meta emphasize solution promotion, organic Bluesky posts emphasize critique, and an LLM-augmented clustering pipeline produces themes that outperform LDA/BERTopic in human evaluation.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Worth engaging for the new Meta–Bluesky dataset and the plausible paid-vs-organic contrast, but the prevalence and event-window numbers rest on mismatched sampling frames and shouldn't be cited without revision. the 4 major comments →

arxiv 2601.13317 v2 pith:SIZTZQZB submitted 2026-01-19 cs.CL cs.AIcs.CYcs.LGcs.SI

Paid Voices vs. Public Feeds: Interpretable Cross-Platform Theme-Based Analysis of Climate Discourse

Samantha Sudhoff , Pranav Perumal , Zhaoqing Wu , Tunazzina Islam This is my paper
classification cs.CL cs.AIcs.CYcs.LGcs.SI
keywords climate discoursecross-platform analysistheme discoveryLLM labelingMeta Ad LibraryBlueskystance predictionevent analysis
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the thematic structure of climate discourse online is systematically shaped by the incentive structure of the platform where it appears. Using a new seed-free theme discovery pipeline, it compares 17,026 paid Meta ads with 19,182 public Bluesky posts from July 2024 to September 2025. It finds that ads emphasize solution promotion—clean energy, solar, infrastructure—while posts emphasize systemic critique—climate rollbacks, science under attack, capitalism's climate doom. It also shows that pro-climate ad themes contracted sharply after the November 2024 U.S. election, while Bluesky themes surged rapidly in response to political events. The paper argues these differences reflect platform-level incentives, not individual intent, and validates the discovered themes through stance prediction and retrieval tasks.

Core claim

The central claim is that platform incentives are reflected in the thematic structure, stance alignment, and temporal responsiveness of climate narratives. Paid advertising on Meta centers on strategic promotion of specific solutions in a formal, forward-looking register, while organic posts on Bluesky center on systemic critique in a crisis-oriented, scientifically grounded register. This divergence is systematic across 51 Meta themes and 46 Bluesky themes, with promotion-oriented themes like 'Clean energy' more frequent in ads and critique-oriented themes like 'Anti-Deforestation' and 'Climate Wildfires' more frequent in posts. Temporal analysis shows that after the November 2024 U.S. elec

What carries the argument

The load-bearing method is a seed-free thematic discovery pipeline: texts are embedded with SentenceBERT, projected with PCA and UMAP, and clustered with HDBSCAN. An LLM (Mistral-Large-Instruct-2407) then filters incoherent clusters, summarizes each cluster, merges redundant clusters based on summary similarity, and assigns concise 1–3 word theme labels. This decouples semantic discovery from theme interpretation, allowing cross-platform comparison without a predefined topic inventory or seed set.

Load-bearing premise

The load-bearing premise is that the two keyword-matched samples measure the same underlying thing called 'climate discourse,' so that differences in theme counts can be attributed to platform incentives rather than to differences in keyword coverage, sampling rates, or deduplication choices.

What would settle it

Compare theme prevalences using identical keyword lists, matched sampling rates, and a range of deduplication thresholds: if the 'Clean energy' gap between Meta and Bluesky narrows or disappears, or if the post-election contraction of pro-climate ads reverses under any plausible reweighting, the central platform-incentive claim would be undermined.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If correct, cross-platform comparisons of climate discourse can be conducted without predefined topic inventories, enabling direct comparison of paid and organic communication across platforms.
  • Pro-climate ad themes contracted sharply after the November 2024 U.S. election, suggesting that climate advertising volume is tied to electoral cycles and political opportunities.
  • Bluesky themes like 'Climate rollbacks' and 'Science under attack' surged immediately after political events, indicating rapid activation of latent frames in organic discourse.
  • The discovered themes improve stance prediction and semantic retrieval performance, suggesting they capture meaningful semantic structure beyond surface-level clusters.
  • The framework's seed-free design could generalize to comparative thematic analysis in other domains and heterogeneous communication environments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A direct testable extension: apply the same pipeline to other ad libraries (e.g., Google Ads) and other organic platforms (e.g., Mastodon) to see whether the paid/organic thematic divide generalizes or is specific to Meta and Bluesky.
  • The post-election contraction of pro-climate ad themes may reflect budget reallocation rather than strategic retreat; a causal design that tracks advertiser budgets or matched ad categories could separate these explanations.
  • The Bluesky stance imbalance (only 4 Pro-Energy out of 500 annotated posts) limits the interpretability of theme-stance correlations there; the paper acknowledges this, but future work could over-sample Pro-Energy posts to stabilize the analysis.
  • The deduplication threshold (cosine similarity 0.80) could inflate cross-platform prevalence differences; re-running the analysis with multiple thresholds would test the robustness of the 'Clean energy' vs. 'Climate Wildfires' gap.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper presents a comparative, corpus-level analysis of climate discourse on two structurally different platforms: paid advertisements from the Meta Ad Library (113,493 keyword-matched ads, deduplicated to 17,026) and organic posts from Bluesky (a random sample of 20,000 from ~1.3M keyword-matched posts, deduplicated to 19,182), covering July 2024–September 2025. The authors introduce an interpretable thematic discovery pipeline: SBERT embeddings, PCA/UMAP, HDBSCAN clustering, LLM-based coherence filtering, summarization, cluster merging, and LLM-generated theme labels. Theme assignments are evaluated against LDA and BERTopic using both human and LLM judges, and the themes are further probed via stance prediction and theme-guided retrieval. The main empirical claims are that paid advertising centers on solution-promotion and forward-looking institutional messaging, while Bluesky centers on systemic critique and crisis-oriented discourse; that these differences are reflected in theme prevalence, stance alignment, and temporal responsiveness; and that the induced themes are more coherent and stance-consistent than baseline topic models.

Significance. If the central comparative claims hold, the paper makes a useful contribution to computational social science and NLP: it jointly analyzes paid ads and organic social media discourse, an underexplored comparison, and proposes an interpretable, seed-free thematic discovery pipeline that could generalize to other cross-platform settings. The authors are explicit that the analysis is descriptive rather than causal, and they acknowledge limitations such as LLM bias and the absence of bias mitigation. The release of the dataset and code is a concrete strength, as is the use of human evaluation as the primary signal for theme quality. However, the quantitative support for the headline cross-platform prevalence and event-responsiveness claims is currently weaker than the prose suggests, primarily because the two corpora are constructed under different sampling and deduplication regimes. The paper's descriptive contrasts are plausible, but several load-bearing comparisons need more careful uncertainty quantification and sensitivity analysis before the strongest conclusions are warranted.

major comments (4)
  1. [Section 3; Section 6.3, Fig. 3] The cross-platform prevalence comparison in Fig. 3 rests on non-comparable sampling frames. Meta uses the complete set of 113,493 keyword-matched ads deduplicated to 17,026 (a 6.7x reduction), while Bluesky uses a random sample of 20,000 posts from ~1.3M, deduplicated to 19,182 (a 1.04x reduction). Because Meta ads are often near-duplicate campaign creatives, deduplication collapses paid amplification into one representative text, while Bluesky posts are almost all retained. It is not shown that theme counts after this asymmetric deduplication are monotonic in actual platform activity (impressions/spend for Meta vs. posting volume for Bluesky). Moreover, Bluesky counts are random-sample estimates with no confidence intervals; a theme such as Climate Wildfires (37 vs. 477 in Fig. 3) could shift under sampling. The claim in Sec. 6.3 that 'remaining differences reflect platform-specific inc
  2. [Section 6.4.2, Fig. 6] The event-window analysis for Bluesky is vulnerable to sampling zeros. The Bluesky corpus consists of ~20k randomly sampled posts over ~15 months, so a three-day window contains on average about 130 sampled posts. A theme with even a 1% background prevalence would be expected to appear only ~1.3 times in that window. The statement that four of five themes were 'entirely absent in the three days preceding the election' and that this indicates 'sudden activation of previously latent frames' is not supported without accounting for the sampling rate. Please report raw counts, per-theme rates with Poisson or bootstrap confidence intervals, or conduct the event analysis on the full ~1.3M post corpus instead of the random sample.
  3. [Section 7.1, Table 3] The stance prediction results are used to validate that themes capture 'stance-relevant signal beyond surface text,' but the evaluation is based on only 500 manually annotated posts per platform, with an 80/20 stratified split yielding a test set of roughly 100 posts. On Bluesky, the stance distribution is extremely imbalanced (4 Pro-Energy posts out of 500), so macro-F1 differences such as 0.46 vs. 0.47 or 0.82 vs. 0.82 are likely within noise. No confidence intervals, significance tests, or variance estimates are reported. This weakens the claim that themes consistently improve stance classification, especially on Bluesky. Please provide uncertainty estimates, consider evaluating on a larger annotation set, or temper the claim accordingly.
  4. [App. C.3] The human evaluation is the primary evidence for theme quality, but the paper does not report inter-annotator agreement (e.g., Cohen's kappa or Krippendorff's alpha). The text states that four researchers annotated in three sessions, but gives no per-annotator breakdown or agreement statistic. Without this, it is difficult to assess whether the reported accuracy differences between methods reflect semantic quality or annotator idiosyncrasy. Please add agreement measures and a more detailed annotation protocol.
minor comments (6)
  1. [Sec. 6.3] Typo: 'Sustainable Fasion' should be 'Sustainable Fashion'. Also, '8 to-tal' appears to be a typo.
  2. [Sec. 5] Typo: 'Principle Component Analysis' should be 'Principal Component Analysis'.
  3. [App. C.2] The sentence 'our Text-to-Theme method outperforms Text-to-Theme' should presumably read 'outperforms Text-to-Summary'.
  4. [Table 3] The table header 'ACCF1' is ambiguous; please label columns as Accuracy and Macro F1.
  5. [Fig. 6] The figure shows top themes but does not indicate whether the bars are counts, proportions, or something else. Please clarify the y-axis and add counts or confidence intervals.
  6. [Section 5.1] For LDA, the number of topics and hyperparameter settings are not reported; please provide these details for reproducibility.

Circularity Check

0 steps flagged

No significant circularity; the thematic pipeline is data-driven and the central claims are empirically grounded rather than reduced to their inputs.

full rationale

The paper's derivation chain is self-contained: themes are induced from raw texts via HDBSCAN semantic clustering followed by LLM-generated labels, with no predefined topic inventory or seed set. The central cross-platform prevalence claims (Fig. 3, Sec. 6.3) are empirical comparisons of LLM-assigned themes across Meta ads and Bluesky posts, not quantities that are equal to any fitted parameter by construction. The LLM-judge evaluation of LLM-produced labels is a potential bias, but the paper explicitly acknowledges it and treats human annotation as the primary signal (App. C.1: 'we treat human annotation as the primary and more reliable evaluation signal'), and downstream stance prediction and retrieval tasks provide independent behavioral validation. Self-citations, including the keyword list from Islam et al. (2023b), are used as data-collection inputs or related-work context rather than as load-bearing uniqueness theorems or ansatz justifications; the keyword set does not define the thematic conclusions. Sampling/dedup asymmetries between the Meta and Bluesky corpora raise validity concerns about prevalence comparisons, but that is a measurement and sampling issue, not a circular reduction of the kind required for a circularity finding. No equation in the paper reduces to its own inputs, and no fitted parameter is renamed as a prediction.

Axiom & Free-Parameter Ledger

6 free parameters · 5 axioms · 0 invented entities

The paper's central comparison depends on a chain of modeling choices and proxy assumptions: keyword-based inclusion, platform-as-proxy, embedding/clustering hyperparameters, and LLM label faithfulness. None of these are derived from first principles, and several are selected on the same data used for evaluation. No new physical or conceptual entities are introduced beyond the discovered theme labels, which are validated only indirectly.

free parameters (6)
  • Deduplication cosine threshold = 0.80
    Used to eliminate highly similar ads/posts; changes the size and composition of both datasets (Section 3).
  • HDBSCAN min_cluster_size / min_samples = 20 / 5
    Clustering hyperparameters set without ablation; directly determine the initial cluster structure (App. B).
  • Merge threshold tau = 0.8
    Selected by grid search on the same data using Silhouette Score and Davies-Bouldin Index; affects the final theme set (App. B.2).
  • Top-k representative texts per cluster = 5
    The LLM sees only the top 5 texts per cluster for coherence and summarization; a different k would change themes (Section 4.1).
  • PCA components / UMAP dimensions = 100 / 20
    Dimensionality reduction choices that affect clustering stability and separability (App. B).
  • Stance classifier hyperparameters (ngram range, C) = e.g., ngram (1,1)-(3,4), C in {0.01..100}
    Tuned on validation set per platform; contributes to the stance prediction results in Table 3 (App. F.1).
axioms (5)
  • domain assumption Keyword-list collection defines the universe of 'climate discourse' on both platforms.
    The study includes only texts matching the curated keywords in Table 5; if keywords systematically favor certain topics, the comparison inherits that bias (Section 3).
  • domain assumption Meta ads are treated as proxies for institutional/resource-backed messaging; Bluesky posts as organic/user-driven discourse.
    Sections 1 and 3 assert structural incentive differences; they are not measured per ad or per post.
  • domain assumption SBERT cosine similarity and HDBSCAN clustering recover meaningful thematic structure in both corpora.
    Clustering quality is only indirectly validated through 500-sample theme accuracy and downstream tasks; no direct cluster-quality ground truth exists (Section 4.1).
  • ad hoc to paper Mistral-Large-Instruct-2407 generates coherent summaries and theme labels that are faithful to the texts.
    The pipeline relies on an LLM for coherence, summary, label, and assignment; no fine-tuning and no bias measurement (acknowledged in Limitations).
  • domain assumption A random sample of 20,000 Bluesky posts is representative of the 1.3M keyword-matched posts.
    Random sampling is stated but no bootstrap, confidence intervals, or demographic comparison are provided (Section 3).

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Paid Voices vs. Public Feeds: Interpretable Cross-Platform Theme-Based Analysis of Climate Discourse." pith.science (2026). https://pith.science/paper/SIZTZQZB

@misc{pith2026260113317,
  author       = {Pith},
  title        = {Pith review of: Paid Voices vs. Public Feeds: Interpretable Cross-Platform Theme-Based Analysis of Climate Discourse},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SIZTZQZB}},
  note         = {Machine review of arXiv:2601.13317}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Climate discourse online shapes public understanding of climate change and informs political and policy debate, yet it unfolds across structurally different environments: paid advertising platforms host targeted, institutionally produced messaging, while public social media reflects largely organic, user-driven discussion. We present a comparative analysis of climate discourse across paid advertisements on Meta (previously Facebook) and public posts on Bluesky from July 2024 to September 2025. To support it, we develop an interpretable thematic discovery pipeline that clusters texts by semantic similarity and uses large language models (LLMs) to label clusters with concise, human-interpretable themes, requiring no predefined topic inventory or seed set. Using these themes, we find the two environments diverge systematically: paid advertising centers on strategic promotion of specific solutions in a formal, forward-looking register, whereas organic discourse centers on systemic critique in a crisis-oriented, scientifically grounded one. We also evaluate the utility of the discovered themes through downstream stance prediction and theme-guided retrieval tasks. While our analysis focuses on climate communication, the framework generalizes to comparative thematic analysis across heterogeneous communication environments.

Figures

Figures reproduced from arXiv: 2601.13317 by Pranav Perumal, Samantha Sudhoff, Tunazzina Islam, Zhaoqing Wu.

Figure 1
Figure 1. Figure 1: Examples of climate discourse across paid [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our framework. social interaction, and grassroots mobilization. Re￾cent work on emerging decentralized platforms such as Bluesky suggests that their design and user migration patterns produce distinct discourse dy￾namics compared to centralized platforms (Quelle and Bovet, 2024). In this work, we treat paid adver￾tising platforms as proxies for institutionally pro￾duced, resource-backed messagi… view at source ↗
Figure 3
Figure 3. Figure 3: # of posts and ads for each common theme. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: UMAP representation of Meta ads and Bluesky posts, annotated with themes. (ex. Meta theme of ‘ Ethical fashion advocacy’ and Bluesky theme of ‘Sustainable luxury fashion’ combine to become, simply, ‘Sustainable Fasion’). A full list of unified themes along with a repre￾sentative example from both Meta and Bluesky is found in [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Thematic shifts on Bluesky surrounding major political events. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Coherency prompt with example response from LLM. Upon removing incoherent examples, we prompt Mistral-Large-Instruct-2407 with the following: "Summarize the following texts by writing 1-3 sentences (<= 100 words) capturing the common theme and intent," shown in [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Prompt used for generating cluster summaries. [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Prompt used for generating cluster themes [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 11
Figure 11. Figure 11: Prompt to assign each text to its correspond [PITH_FULL_IMAGE:figures/full_fig_p015_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Prompt for LLM-based judge to anno￾tate results based on the correctness of assigned key￾words/theme. mediate and implicit summary-to-theme mapping that is influenced by large context windows dur￾ing summary assignment. This design can degrade assignment accuracy, as shown by the LLM and human evaluation results for Bluesky, as well as for the human annotation results for Meta. How￾ever, the Text-to-Summa… view at source ↗
Figure 13
Figure 13. Figure 13: Top 10 themes by number of ads for Meta data. themes, as well as text → theme mapping, are cor￾rect or not for each dataset. In total, it took ≈ 8 hours for both datasets. Annotators were provided with the same instructions that were provided to the LLM prompt. The annotators included advanced undergraduate and graduate students as well as fac￾ulty. D Common Themes in Both Platforms A full list of unified… view at source ↗
Figure 14
Figure 14. Figure 14: Distribution of themes across Meta dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Advertising activity and thematic distribution before and after the November 2024 U.S. election: (a) total [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Advertising activity before and after Charlie Kirk’s death: (a) total spend, (b) total number of ads, and (c) [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Top 10 themes by number of posts for Bluesky data. E.2 Qualitative Analysis on Bluesky The distribution of themes across the Bluesky dataset exhibits a pronounced long-tailed structure ( [PITH_FULL_IMAGE:figures/full_fig_p021_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Distribution of themes across Bluesky dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Correlations between themes and stances for Bluesky. [PITH_FULL_IMAGE:figures/full_fig_p022_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Prompts for downstream task: stance prediction. [PITH_FULL_IMAGE:figures/full_fig_p024_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: t-SNE projection of post and theme embeddings. Themes (blue [PITH_FULL_IMAGE:figures/full_fig_p026_21.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

48 extracted references · 1 canonical work pages

  1. [1]

    Sofiane Abbar, Tahar Zanouda, Laure Berti-Equille, and Javier Borge-Holthoefer. 2016. Using twitter to understand public interest in climate change: The case of qatar. In Proceedings of the International AAAI Conference on Web and Social Media, volume 10, pages 168--177

    2016

  2. [2]

    Gonz \'a lez, and 1 others

    Andreu Arenas, Josep Domingo-Ferrer, Marta C. Gonz \'a lez, and 1 others. 2025. https://doi.org/10.1038/s41597-025-05228-w Comparable 2022 general election advertising datasets from meta and google . Scientific Data, 12(1):1--15

    work page doi:10.1038/s41597-025-05228-w 2025

  3. [3]

    Shraey Bhatia, Jey Han Lau, and Timothy Baldwin. 2021. Automatic classification of neutralization techniques in the narrative of climate change scepticism. In Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 2167--2175

    2021

  4. [4]

    David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. the Journal of machine Learning research, 3:993--1022

    2003

  5. [5]

    Su Lin Blodgett, Solon Barocas, Hal Daum \'e III, and Hanna Wallach. 2020. Language (technology) is power: A critical survey of “bias” in nlp. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454--5476

    2020

  6. [6]

    Emma Frances Bloomfield and Denise Tillery. 2019. The circulation of climate change denial online: Rhetorical and networking strategies on facebook. Environmental Communication, 13(1):23--34

    2019

  7. [7]

    Bluesky Public Social Network . 2025. Bluesky social platform. https://bsky.app/

    2025

  8. [8]

    Alexander Brady and Tunazzina Islam. 2025. Latent topic synthesis: Leveraging llms for electoral ad analysis. arXiv preprint arXiv:2510.15125

    arXiv 2025

  9. [9]

    Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, and 1 others. 2020. Language models are few-shot learners. Advances in neural information processing systems, 33:1877--1901

    2020

  10. [10]

    Arthur Capozzi, Gianmarco De Francisci Morales, Yelena Mejova, Corrado Monti, Andr \'e Panisson, and Daniela Paolotti. 2021. Clandestino or rifugiato? anti-immigration facebook ad targeting in italy. In CHI

    2021

  11. [11]

    David R Cox. 1958. The regression analysis of binary sequences. Journal of the Royal Statistical Society: Series B (Methodological), 20(2):215--232

    1958

  12. [12]

    Biraj Dahal, Sathish AP Kumar, and Zhenlong Li. 2019. Topic modeling and sentiment analysis of global climate change tweets. Social network analysis and mining, 9(1):24

    2019

  13. [13]

    David L Davies and Donald W Bouldin. 2009. A cluster separation measure. IEEE transactions on pattern analysis and machine intelligence, (2):224--227

    2009

  14. [14]

    Maarten Grootendorst. 2022. Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv preprint arXiv:2203.05794

    Pith/arXiv arXiv 2022

  15. [15]

    Kuhika Gupta, Joseph Ripberger, and Wesley Wehde. 2018. Advocacy group messaging on social media: Using the narrative policy framework to study twitter messages about nuclear energy policy in the united states. Policy studies journal, 46(1):119--136

    2018

  16. [16]

    Faye Holder, Sanober Mirza, Jake Carbone, and Ruth E McKie. 2023. Climate obstruction and facebook advertising: how a sample of climate obstruction organizations use social media to disseminate discourses of delay. Climatic Change, 176(2):16

    2023

  17. [17]

    Tunazzina Islam and Dan Goldwasser. 2022. Understanding covid-19 vaccine campaign on facebook using minimal supervision. In 2022 IEEE International Conference on Big Data (Big Data), pages 585--595. IEEE

    2022

  18. [18]

    Tunazzina Islam and Dan Goldwasser. 2025 a . Discovering latent themes in social media messaging: A machine-in-the-loop approach integrating llms. In Proceedings of the International AAAI Conference on Web and Social Media, volume 19, pages 859--884

    2025

  19. [19]

    Tunazzina Islam and Dan Goldwasser. 2025 b . Post-hoc study of climate microtargeting on social media ads with llms: Thematic insights and fairness evaluation. In Findings of the Association for Computational Linguistics: EMNLP 2025

    2025

  20. [20]

    Tunazzina Islam and Dan Goldwasser. 2025 c . Uncovering latent arguments in social media messaging by employing llms-in-the-loop strategy. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 7397--7429

    2025

  21. [21]

    Tunazzina Islam, Shamik Roy, and Dan Goldwasser. 2023 a . Weakly supervised learning for analyzing political campaigns on facebook. In Proceedings of the International AAAI Conference on Web and Social Media, volume 17, pages 411--422

    2023

  22. [22]

    Tunazzina Islam, Ruqi Zhang, and Dan Goldwasser. 2023 b . https://doi.org/10.1145/3600211.3604665 Analysis of climate campaigns on social media using bayesian model averaging . In Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, AIES '23, page 15–25, New York, NY, USA. Association for Computing Machinery

    arXiv 2023

  23. [23]

    Michelle S Lam, Janice Teoh, James A Landay, Jeffrey Heer, and Michael S Bernstein. 2024. Concept induction: Analyzing unstructured text with high-level concepts using lloom. In Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems, pages 1--28

    2024

  24. [24]

    Raul P Lejano and Shondel J Nero. 2020. The power of narrative: Climate skepticism and the deconstruction of science. Oxford University Press

    2020

  25. [25]

    Qiaoyi Liu, Yuheun Kim, and Jeff Hemsley. 2025. Scientists, but deny science? climate change sceptics networks on youtube led by scientists. Information Research an international electronic journal, 30(iConf):741--751

    2025

  26. [26]

    Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    Pith/arXiv arXiv 2019

  27. [27]

    Yiwei Luo, Dallas Card, and Dan Jurafsky. 2020. https://doi.org/10.18653/v1/2020.findings-emnlp.296 Detecting stance in media on global warming . In Findings of the Association for Computational Linguistics: EMNLP 2020, pages 3296--3315, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.findings-emnlp.296 2020

  28. [28]

    Leland McInnes, John Healy, and Steve Astels. 2017. hdbscan: Hierarchical density based clustering. The Journal of Open Source Software, 2(11):205

    2017

  29. [29]

    Leland McInnes, John Healy, and James Melville. 2018. Umap: Uniform manifold approximation and projection for dimension reduction. JOSS

    2018

  30. [30]

    Meta Platforms, Inc. 2025. Meta ad library. https://www.facebook.com/ads/library/

    2025

  31. [31]

    Hopkins, Dan Roth, and Dan Goldwasser

    Nishanth Nakshatri, Nikhil Mehta, Siyi Liu, Sihao Chen, Daniel J. Hopkins, Dan Roth, and Dan Goldwasser. 2025. https://arxiv.org/abs/2504.07400 Talking point based ideological discourse analysis in news events . Preprint, arXiv:2504.07400

    Pith/arXiv arXiv 2025

  32. [32]

    Warren Pearce, Sabine Niederer, Suay Melisa \"O zkula, and Natalia S \'a nchez Querub \' n. 2019. The social media life of climate change: Platforms, publics, and future imaginaries. Wiley interdisciplinary reviews: Climate change, 10(2):e569

    2019

  33. [33]

    Arianna Pera and Luca Maria Aiello. 2024. Shifting climates: Climate change communication from youtube to tiktok. In Proceedings of the 16th ACM Web Science Conference, pages 376--381

    2024

  34. [34]

    Dorian Quelle and Alexandre Bovet. 2024. https://arxiv.org/abs/2405.17571 Bluesky: Network topology, polarization, and algorithmic curation . CoRR

    Pith/arXiv arXiv 2024

  35. [35]

    Nils Reimers and Iryna Gurevych. 2019. https://arxiv.org/abs/1908.10084 Sentence-bert: Sentence embeddings using siamese bert-networks . In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics

    Pith/arXiv arXiv 2019

  36. [36]

    Filipe N Ribeiro, Koustuv Saha, Mahmoudreza Babaei, Lucas Henrique, Johnnatan Messias, Fabricio Benevenuto, Oana Goga, Krishna P Gummadi, and Elissa M Redmiles. 2019. On microtargeting socially divisive ads: A case study of russia-linked ad campaigns on facebook. In ACM FAccT

    2019

  37. [37]

    Peter J Rousseeuw. 1987. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20:53--65

    1987

  38. [38]

    Harri Rowlands, Gaku Morio, Dylan Tanner, and Christopher D Manning. 2024. Predicting narratives of climate obstruction in social media advertising. In Findings of the Association for Computational Linguistics ACL 2024, pages 5547--5558

    2024

  39. [39]

    Kimi Team, Yifan Bai, Yiping Bao, Guanduo Chen, Jiahao Chen, Ningxin Chen, Ruijue Chen, Yanru Chen, Yuankun Chen, Yutian Chen, and 1 others. 2025. Kimi k2: Open agentic intelligence. arXiv preprint arXiv:2507.20534

    Pith/arXiv arXiv 2025

  40. [40]

    Qwen Team. 2025. https://arxiv.org/abs/2505.09388 Qwen3 technical report . Preprint, arXiv:2505.09388

    Pith/arXiv arXiv 2025

  41. [41]

    Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timoth \'e e Lacroix, Baptiste Rozi \`e re, Naman Goyal, Eric Hambro, Faisal Azhar, and 1 others. 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

    Pith/arXiv arXiv 2023

  42. [42]

    Kathie M d'I Treen, Hywel TP Williams, and Saffron J O'Neill. 2020. Online misinformation about climate change. Wiley Interdisciplinary Reviews: Climate Change, 11(5):e665

    2020

  43. [43]

    Apoorva Upadhyaya, Marco Fisichella, and Wolfgang Nejdl. 2023. Intensity-valued emotions help stance detection of climate change twitter data. In IJCAI, pages 6246--6254

    2023

  44. [44]

    Laurens van der Maaten and Geoffrey Hinton. 2008. http://www.jmlr.org/papers/v9/vandermaaten08a.html Visualizing data using t-SNE . Journal of Machine Learning Research, 9:2579--2605

    2008

  45. [45]

    Maryline Vivion, Val \'e rie Trottier, \`E ve Bouh \^e lier, Isabelle Goupil-Sormany, Thierno Diallo, and 1 others. 2024. Misinformation about climate change and related environmental events on social media: Protocol for a scoping review. JMIR Research Protocols, 13(1):e59345

    2024

  46. [46]

    Stefanie Walter, Michael Br \"u ggemann, and Sven Engesser. 2018. Echo chambers of denial: Explaining user comments on climate change. Environmental Communication, 12(2):204--217

    2018

  47. [47]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  48. [48]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.