Pith. sign in

REVIEW 4 major objections 5 minor 38 references

Unveiling Behavioral Differences in Bilingual Information Operations: A Network-Based Approach

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read This paper claims that information-operation drivers on Twitter/X show different topics, media, and coordination patterns in English versus Spanish, so detection must be tuned per language.

desk verdict Honest and clearly written, but the central claim about English vs Spanish IO drivers rests on an unvalidated centrality label that the paper's own appendix undermines. read the letter →

arxiv 2501.09027 v1 pith:F5MH5DIK submitted 2025-01-08 cs.SI

Bowen Yi This is my paper · ORCID
classification cs.SI
keywords informationoperationsTwitter/XSpanish-languagesocialmedianetworkanalysiseigenvectorcentralitybilingualusers2024U.S.electionunsupervisedclustering
verification ladder T0 review T1 audit T2 compute T3 formal

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 information-operation (IO) drivers on Twitter/X behave differently depending on whether they post in English or Spanish, and that a single language-blind detection recipe therefore misses or misreads coordinated influence. Working from a balanced sample of 588,839 tweets per language about the 2024 U.S. presidential election, it builds user-similarity networks from shared domains, shared hashtags, and text similarity, merges them into one fused network, and labels the highest-eigenvector-centrality users in each language as IO drivers. On that basis it reports that English IO drivers concentrate on U.S. politics and right-leaning lower-factuality media, while Spanish IO drivers tilt toward international affairs and mainstream news outlets, with similar engagement levels. It also finds that bilingual posters shift topics and sentiment by language and that many detected English IO drivers, but none of the Spanish ones, were bilingual, which the author uses to warn against treating bilingual status as a coordination signal. The author explicitly flags the absence of labeled data and the unvalidated assumption that high-centrality nodes are IO drivers as limitations.

What carries the argument

The central object is a fused user-similarity network built from three behavioral traces: shared web domains (TF-IDF cosine similarity), shared hashtag sequences, and semantic tweet similarity from multilingual sentence embeddings. On that network the paper applies Louvain community detection, tests network dismantling strategies (weight-, time-, and sentiment-based edge filtering plus eigenvector-centrality node pruning), and then operationally defines IO drivers as the most central users: the top 3% of English nodes and the top 5% of Spanish nodes after filtering. Because no ground-truth labels exist, clustering quality is scored by normalized topic entropy within clusters and Jensen-Shannon divergence between clusters, and the top of the centrality distribution is used as the proxy for coordinated actors.

What would settle it

A labeled set of coordinated and organic accounts from the same 2024 election period would settle the issue: if the top 3% of English nodes and top 5% of Spanish nodes by eigenvector centrality overlap mostly with organic users rather than known IO accounts, then the reported behavioral differences are artifacts of the centrality proxy, not evidence of language-specific information operations.

Watch

Extended reading notes

Core claim

On the paper's own terms, applying an unsupervised, network-based method to balanced English and Spanish Twitter data about the 2024 U.S. election reveals significant behavioral differences between IO drivers in the two languages. English IO drivers predominantly push U.S.-centric hashtags and right-leaning or lower-factuality domains such as breitbart.com and foxnews.com, while Spanish IO drivers favor general-news hashtags and mainstream outlets such as infobae.com and cnn.it, with more attention to international figures and issues like immigration and drug trafficking. The paper also reports that bilingual posters use different topics and sentiment in each language, receive different engagement, and overlap unevenly with the detected IO-driver sets: 53 of 72 English IO drivers were bilingual, whereas none of the 61 Spanish IO drivers were. Because the detection rests on unlabeled centrality, the author states the assumption that high-centrality nodes represent IO drivers and acknowledges it may lead to inaccuracies.

Load-bearing premise

The entire comparison rests on the assumption that the most central users in the fused network are the coordinated IO drivers; if that labeling is wrong, every reported behavioral difference describes central users, not coordinated actors.

Editorial extensions

If this is right

  • If the central claim is correct, IO detection systems tuned on English data should not be transferred directly to Spanish data, because filters and thresholds that work for one language change clustering quality in the other.
  • English and Spanish IO drivers can be distinguished by what they amplify: English drivers concentrate on U.S. partisan hashtags and right-leaning outlets, while Spanish drivers favor breaking-news hashtags and mainstream, often left-of-center, outlets.
  • The finding that most detected English IO drivers were bilingual while no detected Spanish IO driver was bilingual implies that bilingual posting alone is not a reliable indicator of coordinated inauthentic activity.
  • The clustering-quality results, in which none of the fused networks achieved both high intra-cluster homogeneity and high inter-cluster separation, imply that edge filtering and node pruning need per-language optimization rather than a one-size-fits-all recipe.
  • Both English and Spanish IO drivers show similar sentiment trajectories, with positive sentiment spiking around July and August 2024, so the behavioral differences are primarily in topics and media choices, not in emotional valence.
  • If the centrality-based identification is wrong, all of these differences describe central users in general, not coordinated actors, and the paper's own appendix observation that many high-centrality nodes are organic users signals this risk.
  • A direct labeled-data test would be to rerun the pipeline on a dataset with known coordinated accounts and measure how many of the top 3% and top 5% centrality nodes are actually IO drivers, something the author notes is missing.
  • The paper's evidence that bilingual status overlaps unevenly with IO-driver status suggests that a supervised detector using bilingualism as a feature could inherit that unevenness and produce different error rates across languages.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An extension the paper leaves implicit is that the reported behavioral differences are properties of whatever users sit at the top of the centrality distribution, and the appendix's own Botometer spot-checks suggest many of those users may be organic; a labeled evaluation would determine whether the differences describe IO actors or merely central users.
  • The paper's asymmetry, in which 53 of 72 English IO drivers were bilingual while none of the Spanish IO drivers were, could be turned into a testable hypothesis: bilingualism may act as a coordination signal only in the majority-language community, and using it as a feature in a multilingual detector would require per-language calibration.
  • A broader consequence for adjacent work is that unsupervised network methods need an external validity check for what centrality means in each language; replacing fixed percentile cutoffs with a per-language calibration step could make cross-language comparisons more trustworthy.
  • The paper's finding that temporal and sentiment-based edge filtering produced dense, poorly clustered networks suggests that time-window and sentiment thresholds may need to be derived from the data rather than chosen conservatively, which is a direction the paper itself lists as future work.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper analyzes over a million English and Spanish tweets about the 2024 U.S. presidential election, using unsupervised network methods to identify "IO drivers" as high-eigenvector-centrality nodes in fused similarity networks constructed from shared domains, hashtags, and text similarity. It proposes a clustering-quality evaluation based on topic homogeneity and inter-cluster separation, compares network dismantling strategies, and reports behavioral differences between English and Spanish IO drivers in topics, domains, and engagement, as well as analyses of bilingual versus monolingual posters. The authors explicitly acknowledge the lack of labeled data and the assumption that high-centrality nodes represent IO drivers.

Significance. If the central claims were established, the paper would contribute to multilingual IO detection by showing that network-based methods need language-specific tuning and that IO driver behavior differs across languages. The release of code and data, the balanced English-Spanish dataset, and the attempt to evaluate clustering quality without ground truth are strengths. However, the main conclusion that English and Spanish IO drivers exhibit significant behavioral differences currently rests on an unvalidated centrality-based proxy that the paper's own appendix calls into question, and on language-specific pipeline choices that confound the comparison. The paper is a preliminary study, and its findings are better framed as hypotheses than as established results.

major comments (4)
  1. [Section 4.2 and Appendix] The identification of IO drivers as the top 3% (English) and top 5% (Spanish) of nodes by eigenvector centrality is not validated against any ground truth. The Appendix 'Interpretation of High Eigenvector Centrality' explicitly reports that most high-centrality nodes checked corresponded to organic users with low Botometer scores. This directly undermines the premise that high centrality indicates IO activity. Consequently, Tables 2-4 and the topic/sentiment analyses describe high-centrality users rather than coordinated actors, and the claim of 'significant behavioral differences' between English and Spanish IO drivers (Section 5) is not supported. The authors should either incorporate an external validation set (e.g., known IO accounts or coordinated behavior labels) or substantially soften the claims and reframe the analysis as a study of high-centrality users.
  2. [Section 4.1.5 and Section 4.2] The clustering quality evaluation reported in Section 4.1.5 states that 'none of the fused networks achieved satisfactory clustering quality' (high weighted entropy, low JSD). Yet the IO drivers are then selected from these same fused networks, and the choice of the 'best-performing technique' (no filtering for English, weight-based filtering for Spanish) is justified by this internally defined metric. Since the paper itself notes that clustering quality 'does not directly correlate with IO detection performance,' the selection of the pipeline is not well motivated. The centrality scores are derived from networks that the paper finds to be poorly clustered, so the resulting 'IO drivers' may not correspond to meaningful coordinated groups. The authors should demonstrate that the centrality-based identification is robust to this poor clustering, e.g., by comparing results across different clustering algorithms or by showing that the selected nodes exhibit coordination signals beyond network centrality.
  3. [Section 4.2 and Section 3] The language comparison confounds language with pipeline choice. English IO drivers are the top 3% of nodes in an unfiltered fused network, while Spanish IO drivers are the top 5% from a network with bottom 70% of low-weight edges removed. Additionally, several network construction thresholds differ by language (e.g., text similarity threshold 0.90 for English vs. 0.95 for Spanish in Section 4.1.3). Thus, any observed English-Spanish differences in topics, domains, or engagement could be artifacts of the different selection rules rather than genuine behavioral differences. The authors should either use identical filtering and selection criteria for both languages, or explicitly show that the reported differences are robust across a range of thresholds and selection methods.
  4. [Section 4.2 and Section 5] The paper repeatedly uses the phrase 'significant behavioral differences' (e.g., in the abstract and Section 5), but no statistical significance tests are reported. Tables 2-4 are descriptive, and the engagement differences in Table 4 are small (e.g., average likes 28.77 vs. 28.37). Without confidence intervals, hypothesis tests, or effect sizes, the claim of statistical significance is unsupported. The authors should add appropriate inferential statistics (e.g., permutation tests, bootstrap CIs) for the comparisons they make.
minor comments (5)
  1. [Section 4.1.5] The entropy formula uses the base of the logarithm implicitly; please specify that the log is natural or base 2, and clarify the normalization by log(|T_k|) when |T_k| is small.
  2. [Section 4.3] The sentence 'Average quotes and quotes can also be found in the code file' appears to have a typo; likely 'Average quotes and replies'.
  3. [Section 5, Limitations] The Limitations state that the high-centrality assumption 'may lead to inaccuracies in supervised settings,' but the paper uses unsupervised settings; the wording should be 'unsupervised' or 'both supervised and unsupervised.'
  4. [Section 4.1.2] The phrase 'user @StrokerAce90 had the maximum centrality across two of the largest clusters' is ambiguous; clarify whether this user had the highest centrality in each of the two clusters or something else.
  5. [References] The reference for Jiang and Xu (2023) appears in the text but is missing from the reference list; please add it.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the IO-driver label is an unvalidated centrality proxy, but the behavioral comparisons are independent descriptive summaries, not reductions to the selection rule.

full rationale

The paper's derivation chain is: behavioral traces (shared domains, hashtags, text) are converted into similarity networks; the networks are fused; eigenvector centrality is computed; the top 3% (English) or top 5% (Spanish) of nodes are labeled IO drivers; and the tweets of those nodes are then summarized by topics, domains, sentiment, and engagement. The only step that could appear circular is the operationalization of 'IO driver' as high eigenvector centrality. However, the subsequent behavioral comparisons are not derived from the centrality equations or from the network-construction formulas; they are empirical summaries of the selected users' content and engagement. The paper is explicit that this is an assumption, stating in the limitations that 'the assumption that high-centrality nodes represent IO drivers may lead to inaccuracies in supervised settings,' and the appendix honestly reports that many high-centrality nodes correspond to organic users with low Botometer scores. Those admissions undermine the external validity of calling the selected nodes IO drivers, but they do not make the claimed behavioral differences equivalent to the selection rule by construction. The language-specific thresholds and edge-filtering choices (3% vs 5%; unfiltered vs weight-filtered) are a confounding design choice, not a circular reduction. The method basis in Luceri et al. (2023) comes from non-overlapping authors and is used as inspiration for network fusion and node pruning, not as a self-citation chain or an imported uniqueness theorem. No equations in the paper reduce the conclusions to their inputs, and no fitted parameter is renamed as a prediction. The central validity risk is the unvalidated centrality proxy, which is a correctness concern, not a circularity finding under the stated criteria.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central analysis rests on hand-set similarity thresholds, centrality-based IO driver labeling, and an unvalidated association between cluster metrics and IO detection quality. No new physical or conceptual entities are introduced.

free parameters (6)
  • Cosine similarity threshold for Co-Domain network = 0.6
    Hand-set threshold for retaining user-user edges; affects network density and all downstream results.
  • Cosine similarity threshold for Co-Hashtag network = 0.7
    Hand-set threshold, chosen without a tuning search or external validation.
  • Text similarity threshold = 0.90 for English, 0.95 for Spanish
    Adjusted per language to obtain comparable network sizes; this choice directly impacts clustering and IO driver detection.
  • Eigenvector centrality pruning threshold = 10^-2
    Used in node pruning experiments; selected based on prior work without data-specific justification.
  • IO driver selection percentiles = English top 3% no filtering; Spanish top 5% after removing bottom 70% edges
    Differing criteria are ad hoc adjustments for graph size differences; they determine who is labeled an IO driver and therefore shape all behavioral comparisons.
  • Temporal alignment window = 1 hour
    Edges kept only if tweets occur within one hour; described as conservative, but no sensitivity analysis is provided.
assumptions (5)
  • domain assumption Louvain community detection with resolution 1.0 yields meaningful clusters for IO detection
    Used throughout without comparison to other clustering algorithms or parameters; acknowledged in Limitations.
  • domain assumption High eigenvector centrality in fused behavioral networks identifies IO drivers
    Explicitly stated in Limitations: 'the assumption that high-centrality nodes represent IO drivers may lead to inaccuracies.' The appendix also reports high-centrality organic users.
  • domain assumption Shared domains, hashtags, and text similarity are valid behavioral traces of coordinated activity
    Borrowed from Luceri et al. 2023 and Pacheco et al. 2020; accepted without independent verification in this dataset.
  • ad hoc to paper Topic homogeneity and inter-cluster JSD are meaningful proxies for clustering quality and IO detection utility
    Proposed by the authors as an unsupervised evaluation; not validated against any ground truth or externally known IO driver set.
  • domain assumption Downsampling English tweets to Spanish size removes confounders
    Authors state 'completely eliminating confounding factors remains challenging' in Section 3, yet the analysis proceeds as if the balance is sufficient.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unveiling Behavioral Differences in Bilingual Information Operations: A Network-Based Approach." pith.science (2026). https://pith.science/paper/F5MH5DIK

@misc{pith2026250109027,
  author       = {Pith},
  title        = {Pith review of: Unveiling Behavioral Differences in Bilingual Information Operations: A Network-Based Approach},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/F5MH5DIK}},
  note         = {Machine review of arXiv:2501.09027}
}
read the original abstract

Twitter has become a pivotal platform for conducting information operations (IOs), particularly during high-stakes political events. In this study, we analyze over a million tweets about the 2024 U.S. presidential election to explore an under-studied area: the behavioral differences of IO drivers from English- and Spanish-speaking communities. Using similarity graphs constructed from behavioral patterns, we identify IO drivers in both languages and evaluate the clustering quality of these graphs in an unsupervised setting. Our analysis demonstrates how different network dismantling strategies, such as node pruning and edge filtering, can impact clustering quality and the identification of coordinated IO drivers. We also reveal significant differences in the topics and political indicators between English and Spanish IO drivers. Additionally, we investigate bilingual users who post in both languages, systematically uncovering their distinct roles and behaviors compared to monolingual users. These findings underscore the importance of robust, culturally and linguistically adaptable IO detection methods to mitigate the risks of influence campaigns on social media. Our code and data are available on GitHub: https://github.com/bowenyi-pierre/humans-lab-hackathon-24.

Figures

Figures reproduced from arXiv: 2501.09027 by the authors.

Figure 1
Figure 1. Comparison of the 5 most popular web domains in English and Spanish tweets. Factuality and political [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Co-Domain networks for English (left) and Spanish (right) tweets. Each node represents a Twitter account, [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Co-Hashtag networks for English (left) and Spanish (right) tweets. Node sizes are proportional to [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Text similarity networks for English (left) and Spanish (right) tweets. Except for common cluster [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Clustering quality evaluation for English (left) and Spanish (right) fused networks after five different [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Sentiment shift of English and Spanish tweets [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Sentiment trend for English IO drivers. Over [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Sentiment trend for Spanish IO drivers. Over [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 24 canonical work pages

  1. [1]

    Raquel Alvarez, David Garcia, Yamir Moreno, and Frank Schweitzer. 2015. Sentiment cascades in the 15m movement. EPJ Data Science, 4:1--13

    work page 2015

  2. [2]

    Adam Badawy, Emilio Ferrara, and Kristina Lerman. 2018. https://doi.org/10.1109/ASONAM.2018.8508646 Analyzing the digital traces of political manipulation: The 2016 russian interference twitter campaign . In 2018 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 258--265

    work page arXiv 2018

  3. [3]

    Ashwin Balasubramanian, Vito Zou, Hitesh Narayana, Christina You, Luca Luceri, and Emilio Ferrara. 2024. https://arxiv.org/abs/2411.00376 A public dataset tracking social media discourse about the 2024 u.s. presidential election on twitter/x . Preprint, arXiv:2411.00376

    arXiv 2024

  4. [4]

    Mathieu Bastian, Sebastien Heymann, and Mathieu Jacomy. 2009. Gephi: an open source software for exploring and manipulating networks. In Proceedings of the international AAAI conference on web and social media, volume 3, pages 361--362

    work page 2009

  5. [5]

    Vincent D Blondel, Jean-Loup Guillaume, Renaud Lambiotte, and Etienne Lefebvre. 2008. Fast unfolding of communities in large networks. Journal of statistical mechanics: theory and experiment, 2008(10):P10008

    work page 2008

  6. [6]

    Paula Bogonez Mu \ n oz. 2023. The russian narrative strategy on telegram. analyzing propaganda, disinformation, and the pursuit of influence in colombia and mexico in the context of ukraine's war

    work page 2023

  7. [7]

    Keith Burghardt, Ashwin Rao, Siyi Guo, Zihao He, Georgios Chochlakis, Baruah Sabyasachee, Andrew Rojecki, Shri Narayanan, and Kristina Lerman. 2023. https://arxiv.org/abs/2305.11867 Socio-linguistic characteristics of coordinated inauthentic accounts . Preprint, arXiv:2305.11867

    work page Pith review arXiv 2023

  8. [8]

    Federico Cinus, Marco Minici, Luca Luceri, and Emilio Ferrara. 2024. https://arxiv.org/abs/2410.22716 Exposing cross-platform coordinated inauthentic activity in the run-up to the 2024 u.s. election . Preprint, arXiv:2410.22716

    work page Pith review arXiv 2024

Show all 38 references
  1. [9]

    Bent Fuglede and Flemming Topsoe. 2004. Jensen-shannon divergence and hilbert space embedding. In International symposium onInformation theory, 2004. ISIT 2004. Proceedings., page 31. IEEE

    2004

  2. [10]

    Gallegos, Ryan A

    Isabel O. Gallegos, Ryan A. Rossi, Joe Barrow, Md Mehrab Tanjim, Sungchul Kim, Franck Dernoncourt, Tong Yu, Ruiyi Zhang, and Nesreen K. Ahmed. 2024. https://arxiv.org/abs/2309.00770 Bias and fairness in large language models: A survey . Preprint, arXiv:2309.00770

    2024 arXiv

  3. [11]

    Ruth Olimpia Garcia Gavilanes, Diego G \'o mez, Denis Parra, Christoph Trattner, Andreas Kaltenbrunner, and Eduardo Graells-Garrido. 2015. https://api.semanticscholar.org/CorpusID:8099267 Language, twitter and academic conferences . Proceedings of the 26th ACM Conference on Hy...

    2015

  4. [12]

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

    2022 arXiv

  5. [13]

    Samar Haider, Luca Luceri, Ashok Deb, Adam Badawy, Nanyun Peng, and Emilio Ferrara. 2020. https://api.semanticscholar.org/CorpusID:226299800 Detecting social media manipulation in low-resource languages . Companion Proceedings of the ACM Web Conference 2023

    2020

  6. [14]

    Lynnette Hui, Xian Ng, J. D. Moffitt, and Kathleen M. Carley. 2022. https://api.semanticscholar.org/CorpusID:249461806 Coordinated through aweb of images: Analysis of image-based influence operations from china, iran, russia, and venezuela . ArXiv, abs/2206.03576

    2022 arXiv

  7. [15]

    Jacobs and Kathleen M

    Charity S. Jacobs and Kathleen M. Carley. 2022. https://doi.org/10.1007/978-3-031-17114-7_9 \#whodefinesdemocracy: Analysis on 2021 chinese messaging campaign . In Social, Cultural, and Behavioral Modeling: 15th International Conference, SBP-BRiMS 2022, Pittsburgh, PA, USA, Se...

    2022 doi

  8. [16]

    Charity S Jacobs and Kathleen M Carley. 2024. \# whatisdemocracy: finding key actors in a chinese influence campaign. Computational and Mathematical Organization Theory, 30(2):127--147

    2024

  9. [17]

    Ke Jiang and Qian Xu. 2023. https://api.semanticscholar.org/CorpusID:264324059 Analyzing the dynamics of social media texts using coherency network analysis: a case study of the tweets with the co-hashtags of \#blacklivesmatter and \#stopasianhate . Frontiers in Research Metri...

    2023

  10. [18]

    Franziska Keller, David Schoch, Sebastian Stier, and JungHwan Yang. 2017. How to manipulate social media: Analyzing political astroturfing using ground truth data from south korea. In Proceedings of the international AAAI conference on web and social media, volume 11, pages 564--567

    2017

  11. [19]

    Franziska B Keller, David Schoch, Sebastian Stier, and JungHwan Yang. 2020. Political astroturfing on twitter: How to coordinate a disinformation campaign. Political communication, 37(2):256--280

    2020

  12. [20]

    Luca Luceri, Valeria Pantè, Keith Burghardt, and Emilio Ferrara. 2023. https://arxiv.org/abs/2310.09884 Unmasking the web of deceit: Uncovering coordinated activity to expose information operations on twitter . Preprint, arXiv:2310.09884

    2023 arXiv

  13. [21]

    Sarah Masud, Sahajpreet Singh, Viktor Hangya, Alexander Fraser, and Tanmoy Chakraborty. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.886 Hate personified: Investigating the role of LLM s in content moderation . In Proceedings of the 2024 Conference on Empirical Methods in...

    2024 doi

  14. [22]

    Julia Mendelsohn, Sayan Ghosh, David Jurgens, and Ceren Budak. 2023. https://api.semanticscholar.org/CorpusID:258048458 Bridging nations: Quantifying the role of multilinguals in communication on social media . In International Conference on Web and Social Media

    2023

  15. [23]

    Marco Minici, Luca Luceri, Federico Cinus, and Emilio Ferrara. 2024. https://api.semanticscholar.org/CorpusID:272832436 Uncovering coordinated cross-platform information operations threatening the integrity of the 2024 u.s. presidential election online discussion . ArXiv, abs/...

    2024 arXiv

  16. [24]

    Diogo Pacheco, Alessandro Flammini, and Filippo Menczer. 2020 a . https://doi.org/10.1145/3366424.3385775 Unveiling coordinated groups behind white helmets disinformation . In Companion Proceedings of the Web Conference 2020, WWW '20, page 611–616, New York, NY, USA. Associati...

    2020

  17. [25]

    Diogo Pacheco, Pik-Mai Hui, Christopher Torres-Lugo, Bao Tran Truong, Alessandro Flammini, and Filippo Menczer. 2020 b . https://api.semanticscholar.org/CorpusID:235428473 Uncovering coordinated networks on social media . ArXiv, abs/2001.05658

    2020 arXiv

  18. [26]

    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

    2019 arXiv

  19. [27]

    Andrew Rosenberg and Julia Hirschberg. 2007. V-measure: A conditional entropy-based external cluster evaluation measure. In Proceedings of the 2007 joint conference on empirical methods in natural language processing and computational natural language learning (EMNLP-CoNLL), p...

    2007

  20. [28]

    Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. https://doi.org/10.18653/v1/P19-1163 The risk of racial bias in hate speech detection . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668--1678, ...

    2019 doi

  21. [29]

    Mohsen Sayyadiharikandeh, Onur Varol, Kai-Cheng Yang, Alessandro Flammini, and Filippo Menczer. 2020. https://doi.org/10.1145/3340531.3412698 Detection of novel social bots by ensembles of specialized classifiers . In Proceedings of the 29th ACM International Conference on Inf...

    2020

  22. [30]

    Kashish Shah, Patrick Gerard, Luca Luceri, and Emilio Ferrara. 2024. https://api.semanticscholar.org/CorpusID:273812425 Unfiltered conversations: A dataset of 2024 u.s. presidential election discourse on truth social . ArXiv, abs/2411.01330

    2024 arXiv

  23. [31]

    Karishma Sharma, Yizhou Zhang, Emilio Ferrara, and Y. Liu. 2020. https://api.semanticscholar.org/CorpusID:234771738 Identifying coordinated accounts on social media through hidden influence and group behaviours . Proceedings of the 27th ACM SIGKDD Conference on Knowledge Disco...

    2020

  24. [32]

    Vishnuprasad Padinjaredath Suresh, Gianluca Nogara, Felipe Cardoso, Stefano Cresci, Silvia Giordano, and Luca Luceri. 2023. https://arxiv.org/abs/2302.04450 Tracking fringe and coordinated activity on twitter leading up to the us capitol attack . Preprint, arXiv:2302.04450

    2023 arXiv

  25. [33]

    Xinyu Wang, Jiayi Li, Eesha Srivatsavaya, and Sarah Rajtmajer. 2023. Evidence of inter-state coordination amongst state-backed information operations. Scientific reports, 13(1):7716

    2023

  26. [34]

    Yan Xia, Ted Hsuan Yun Chen, and Mikko Kivel \"a . 2020. https://api.semanticscholar.org/CorpusID:235747929 Spread of tweets in climate discussions: A case study of the 2019 nobel peace prize announcement . Nordic Journal of Media Studies, 3:96 -- 117

    2020

  27. [35]

    Kai-Cheng Yang, Emilio Ferrara, and Filippo Menczer. 2022. Botometer 101: Social bot practicum for computational social scientists. Journal of computational social science, 5(2):1511--1528

    2022

  28. [36]

    Savvas Zannettou, Tristan Caulfield, William Setzer, Michael Sirivianos, Gianluca Stringhini, and Jeremy Blackburn. 2018. https://api.semanticscholar.org/CorpusID:53213951 Who let the trolls out?: Towards understanding state-sponsored trolls . Proceedings of the 10th ACM Confe...

    2018

  29. [37]

    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...

  30. [38]

    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 gl...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.