CoChat
Sign Up
Cited by 2Social Media and PoliticsMisinformation and Its ImpactsHate Speech and Cyberbullying Detection

Cross-Partisan Interactions on Twitter

Open access
Çetinkaya, Yusuf Mücahit, Ghafouri, Vahid, Suarez-Tangil, Guillermo, Such, Jose, Elmas, Tuğrulcan·2025·Proceedings of the International AAAI Conference on Web and Social Media
DOI 10.1609/icwsm.v19i1.35819
Open this paper in CoChat — free accountOpen PDF

Abstract

Many social media studies argue that social media creates echo chambers where some users only interact with peers of the same political orientation. However, recent studies suggest that a substantial amount of Cross-Partisan Interactions (CPIs) do exist --- even within echo chambers, but they may be toxic. There is no consensus about how such interactions occur and when they lead to healthy or toxic dialogue. In this paper, we study a comprehensive Twitter dataset that consists of 3 million tweets from 2020 related to the U.S. context to understand the dynamics behind CPIs. We investigate factors that are more associated with such interactions, including how users engage in CPIs, which topics are more contentious, and what are the stances associated with healthy interactions. We find that CPIs are significantly influenced by the nature of the topics being discussed, with politically charged events acting as strong catalysts. The political discourse and pre-established political views sway how users participate in CPIs, but the direction in which users go is nuanced. While Democrats engage in cross-partisan interactions slightly more frequently, these interactions often involve more negative and nonconstructive stances compared to their intra-party interactions. In contrast, Republicans tend to maintain a more consistent tone across interactions. Although users are more likely to engage in CPIs with popular accounts in general, this is less common among Republicans who often engage in CPIs with accounts with a low number of followers for personal matters. Our study has implications beyond Twitter as identifying topics with low toxicity and high CPI can help highlight potential opportunities for reducing polarization while topics with high toxicity and low CPI may action targeted interventions when moderating harm.

AI snapshot

Written from the paper’s full text.

TL;DR

This paper analyzes a 2020 U.S. Twitter reply dataset to identify which users, topics, and content features are associated with cross-partisan interactions (CPIs). It finds CPIs are shaped strongly by topic and user attributes, are often more toxic than intra-party replies, but some high-CPI/low-toxicity topics may offer bridge opportunities.

Key findings

  • The final dataset contains 2,561,846 tweets from 726,140 users; 661,661 replies (34%) are classified as CPIs. Democrats originated 65% of CPIs, with 34.3% of Democrat tweets and 33.2% of Republican tweets being cross-partisan.
  • Topic matters: politics/governance and political figures generate the highest CPI levels, with topics such as “Impeachment Proceedings and Senate Actions” and “Election Strategies” at 53–60% CPI, and sports events above 55%; public health mask/hygiene topics are 27–36%, and hobbies/religious topics 23–30%. Topic CPI and toxic content percentages have a Pearson correlation of 0.76.
  • Repliers in CPIs tend to have fewer followers, followings, and tweets, while root users’ attributes matter more. Root reply count has the largest positive coefficient overall (0.525; Democrats 1.978, Republicans -0.304); root like count is negative for Democrats (-1.450) and positive for Republicans (0.788).
  • CPI replies are consistently more toxic than non-CPI replies as political orientation moves away from the center, with lowest toxicity near the center (CPI 3.1%, non-CPI 2.7%). Democrats show more negative/confrontational stances in D→R than D→D, while Republicans maintain a more consistent tone; root sentiment–reply stance patterns correlate 95% between partisan and cross-partisan interactions.

Methods

  • Used the Internet Archive’s 1% Twitter sample, recollecting replies and root tweets via Twitter API in March 2023, limited to English direct replies to root tweets in the 2020 U.S. context. Political orientation was measured with Barberá’s Bayesian ideal-point method (-5 to +5); a CPI is a reply between users with different orientations.
  • Toxicity was scored with Google Jigsaw’s Perspective API, taking the maximum score across categories and using a 0.61 threshold. Topic modeling used BERTopic (BERT embeddings, UMAP, HDBSCAN) on root tweets, identifying 177 topics grouped into 11 categories.
  • LLM annotation used Mistral-7B-Instruct-v0.2 to generate three adjectives for root sentiment and reply stance, sampling 100,000 tweet-reply pairs per interaction type (400,000 total). Validation on 100 pairs gave 97%/88% sentiment accuracy and 88%/85% stance accuracy by two annotators. Logistic and Lasso regressions modeled user/CPI associations.

Limitations

  • Stated: continuous CPI values were not deeply analyzed; analysis is limited to root tweets and direct replies, excluding reply chains and media/links/images/videos; LLM annotation sometimes confused stance and sentiment or generated sentences; findings are specific to X/Twitter and the 2020 U.S. context under a Republican government.
  • Evident from the text: political orientation classification excludes users without following data and neutral users, so results apply only to classified non-neutral partisans; toxicity classification depends on Perspective API and the 0.61 threshold, though sensitivity checks from 0.4 to 0.9 reportedly showed no substantial change.
  • Evident from the text: the regression models have modest explanatory power—full logistic model Pseudo R-squared 0.05413 and ROC AUC 0.656, with the Republican model ROC AUC 0.538—so reported associations are not strong causal predictions.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1609/icwsm.v19i1.35819

Cites

0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

9 of 14 cited sources checked · 6 supported · 1 flagged · accuracy not yet measured

  • Findingown data

    Political topics comprise approximately 35% of the 177 topics.

    “Political topics comprise approximately 35% of these 177 topics”
    Quote in the paperNot a substantial claim
  • Findingown data

    The topic modeling identifies 177 distinct topics within the dataset.

    “The topic modeling identifies 177 distinct topics within the dataset.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Toxicity levels for CPI replies are consistently higher than for non-CPI replies as the political orientation score moves away from the center.

    “The overall trend shows that toxicity levels for CPI replies (red bars) are consistently higher than for non-CPI replies (blue bars) as the political orientation score moves away from the center.”
    Quote in the paperSubstantial claim
  • Findingnone given

    Users are more likely to engage in CPIs with popular accounts in general.

    “Although users are more likely to engage in CPIs with popular accounts in general”
    Quote in the paperNot a substantial claim
  • Findingown data

    There are 661,661 replies classified as CPI (34%).

    “There are 661,661 replies classified as CPI (34%).”
    Quote in the paperSubstantial claim
  • Classified as probably not a substantial claim. 20
  • May combine two claims. 23

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    The toxicity threshold is set at 0.61, identified by Kumar et al. (2021) and Aleksandric et al. (2024).

    “We set the threshold at 0.61 which is identified by Kumar et al. (2021) by employing a survey on 17,280 participants and Aleksandric et al. (2024) who found that Perspective API gets the highest accuracy on this threshold on their independent test set.”
    Quote in the paperSubstantial claim2 of 2 sources checked
    • cites Kumar et al. (2021) source

      Section 6.1 (p. 11) reports 0.61 as the median of thresholds optimized separately for individual participants, not as a recommended single toxicity threshold. It instead gives 0.18–0.49 as the optimal range for the aggregate dataset. The Conclusion (p. 12) confirms that the study surveyed 17,280 participants. (LLM, conf=0.90)

      Experimental — accuracy not yet measured
    • cites Aleksandric et al. (2024) source

      In “GPT Labels” (page not provided), Aleksandric et al. report that Perspective API achieved its highest accuracy at a threshold of 0.6, not 0.61, and say they used 0.6 to classify toxic replies. They compared the API against a manually labeled sample; the passage does not describe an independent test set. (LLM, conf=0.90)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The median optimal threshold for classifying toxic content across participants is 0.61.

    “They find that the median optimal threshold across participants 0.61.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Kumar et al. (2021) source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Santoro and Broockman (2022) find that the effectiveness of cross-partisan conversations depends on the topic through a randomized controlled trial.

    “Santoro and Broockman (2022) find that the effectiveness of cross-partisan conversations depends on the topic through a randomized controlled trial.”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Santoro and Broockman (2022) source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    A substantial portion of the audience accesses news from outlets with varying ideological perspectives from both online and offline platforms.

    “A substantial portion of the audience accesses news from outlets with varying ideological perspectives from both online and offline platforms (Gentzkow and Shapiro 2011).”
    Quote in the paperNot a substantial claim1 of 1 sources checked
    • cites Gentzkow and Shapiro 2011 source

      The Introduction reports that a significant share of consumers get online news from multiple outlets, including outlets with differing ideological audiences. Section 4.1 finds low ideological segregation in several offline media, but it does not establish that a substantial portion of individuals accesses differently oriented outlets across both online and offline platforms. (LLM, conf=0.84)

      Experimental — accuracy not yet measured

  • Classified as probably not a substantial claim.

    Study design and conductcitation

    The methodology of Barberá (2015) measures users' political orientation.

    “We employ the methodology of Barberá (2015) to measure users' political orientation.”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites Barberá 2015 source

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Past studies have not reached a consensus on whether cross-partisan interactions bridge divides or exacerbate tensions.

    “Past studies have not reached a consensus on whether such interactions bridge divides or exacerbate tensions”
    Quote in the paperNot a substantial claim1 of 2 sources checked
    • cites Bail et al. (2018) source
    • cites Rajadesingan et al. (2021)

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    LLMs are capable of text mining tasks such as stance detection, sentiment analysis, hate speech, and sociopolitical affiliation of texts.

    “recent studies suggest that LLMs are capable of text mining tasks such as stance detection (Lan et al. 2024), sentiment analysis (Lee et al. 2024), hate speech (Kaya, Ozcelik, and Toraman 2024), and sociopolitical affiliation of texts (Ghafouri et al. 2023).”
    Quote in the paperNot a substantial claim2 of 3 sources checked
    • cites Lan et al. 2024 source
    • cites Lee et al. 2024

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites Kaya, Ozcelik, and Toraman 2024 source

  • May combine two claims.

    Findingnone given

    Repliers in CPIs tend to have lower number of followers, followings, and tweet counts.

    “In contrast, the repliers in CPIs tend to have lower number of followers, followings and tweet counts.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Root users' tweet counts have negative association with CPIs, while their following counts have contrasting but negligible associations.

    “Root users' tweet counts have negative association while their following counts have contrasting but negligible associations with CPIs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Of all Republican tweets, 33.2% are Republicans replying to Democrats; of all Democrat tweets, 34.3% are Democrats replying to Republicans.

    “Of these, 196,642 are from Republicans replying to Democrats, making up 33.2% of all Republican tweets, and 432,004 are from Democrats replying to Republicans, accounting for 34.3% of all Democrat tweets.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Around 10-15% of topics specifically pertain to election-related discourse.

    “A notable subset of these, around 10-15%, specifically pertains to electionrelated discourse”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The political discourse and pre-established political views sway how users participate in CPIs, but the direction in which users go is nuanced.

    “The political discourse and pre-established political views sway how users participate in CPIs, but the direction in which users go is nuanced.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Accounts fostering CPIs by replying have fewer followers, followings, and tweets (negative coefficients).

    “We observe that accounts fostering CPIs by replying have fewer count of followers, following, and tweets as these have negative coefficients.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Root tweets from users with high follower count are more associated with CPIs from Democrats but less with CPIs from Republicans.

    “On the other hand, root tweets from users with high follower count are more associated with CPIs from Democrats but less with CPIs from Republicans.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    CPIs are significantly influenced by the nature of the topics being discussed, with politically charged events acting as strong catalysts.

    “We find that CPIs are significantly influenced by the nature of the topics being discussed, with politically charged events acting as strong catalysts.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingnone given

    Among Republicans, engaging in CPIs with popular accounts is less common; they often engage in CPIs with accounts with a low number of followers for personal matters.

    “this is less common among Republicans who often engage in CPIs with accounts with a low number of followers for personal matters.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    The tweet's retweet count has a similar association and generalizes to Republicans although the coefficient is low.

    “The tweet's retweet count has a similar association and generalizes to Republicans although the coefficient is low.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Toxicity scores gradually decrease toward the center, reaching their lowest around the center, where CPI toxicity drops to 3.1% and non-CPI toxicity to 2.7%.

    “As moving toward the center, toxicity scores gradually decrease for both CPI and non-CPI interactions, reaching their lowest around the center, where CPI toxicity drops to 3.1% and non-CPI toxicity to 2.7%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Democrats tend not to be involved in CPIs, as their bins have more negative coefficient portions compared to Republicans, who are more likely to engage in CPIs.

    “Our analysis reveals that Democrats tend not to be involved in CPIs, as their bins have more negative coefficient portions compared to Republicans, who are more likely to engage in CPIs.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Like count has a very high negative coefficient for Democrats and a very high positive coefficient for Republicans.

    “Surprisingly, like count has a very high negative coefficient for Democrats and a very high positive coefficient for Republicans.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The rich-get-richer effect is more pronounced for Democrats but does not generalize to Republicans, which has a negative coefficient for the same attribute.

    “This findings is even more pronounced for Democrats, but does not generalize to Republicans, which has a negative coefficient for the same attribute.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Cross-partisan engagement is influenced not only by root tweet attributes but also by the alignment of the repliers.

    “The differences between the camps underline how crosspartisan engagement is influenced not only by root tweet attributes but also by the alignment of the repliers”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationcitation

    The dynamics of cross-partisan interactions are complex, and their impact varies on a case-by-case basis.

    “These studies suggest that the dynamics of cross-partisan interactions are complex, and their impact varies on a caseby-case basis.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationcitation

    Individuals are not wholly isolated within echo chambers but rather encounter diverse viewpoints.

    “Such cross-partisan media consumption indicates that individuals are not wholly isolated within echo chambers but rather encounter diverse viewpoints.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    CPIs can bridge ideological gaps but also risk escalating tensions.

    “which suggests that while these interactions can bridge ideological gaps, they also risk escalating tensions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationown data

    Republicans are more inclined to engage in CPIs with highly liked tweets from the opposing camp, whereas Democrats may tend to create discussions on tweets they dislike.

    “This indicates that Republicans are more inclined to engage in CPIs with highly liked tweets from the opposing camp, whereas Democrats may tend to create discussions on tweets they (and the general public) dislike.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The code is fully shared online.

    “the code is fully shared online”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The data used in the study was opensourced.

    “the data we used was opensourced”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The analysis does not reveal the identity of the users.

    “our analysis does not reveal the identity of the users”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    All training details were specified.

    “Did you specify all the training details (e.g., data splits, hyperparameters, how they were chosen)?Yes”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable). May combine two claims.

    Interpretationreasoning

    These bottom-right quadrant topics offer opportunities for bridging societal divides, acting as contact points where cross-partisan dialogue can flourish.

    “which offer opportunities for bridging societal divides, acting as contact points where cross-partisan dialogue can flourish.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    Politically and socially significant topics tend to disrupt echo chamber dynamics, while more personal or neutral topics may reinforce them, leading to isolated discussions within homogeneous groups.

    “The findings suggest that while politically and socially significant topics tend to disrupt echo chamber dynamics, more personal or neutral topics may reinforce them, leading to isolated discussions within homogeneous groups.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The total amount of compute and the type of resources used are included in the Data & Methodology section.

    “Did you include the total amount of compute and the type of resources used (e.g., type of GPUs, internal cluster, or cloud provider)?Yes, see Data & Methodology section.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductreasoning

    The machine learning experiment was for generating text, so accuracy and error bars do not apply.

    “Our machine learning experiment was done for generating text where the concept of accuracy and error-bar does not apply.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S4 (based). May combine two claims.

    Study design and conductnone given

    The proposed evaluation is justified as sufficient and appropriate to the claims made.

    “Do you justify how the proposed evaluation is sufficient and appropriate to the claims made?Yes”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The labels are shared on GitHub for reproducibility.

    “we share our labels on GitHub for reproducibility”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Study design and conductnone given

    The cost of misclassification and fault tolerance is discussed.

    “Do you discuss what is "the cost" of misclassification and fault (in)tolerance?Yes”
    Quote in the paperNot a substantial claim
37 of 37

  • The paper relies on this without support.

    A case-by-case understanding of the likeliness and impact of CPIs is not clearly established.

    “To the best of our knowledge, such a case-by-case understanding of the likeliness and impact of CPIs is not clearly established.”

  • The paper relies on this without support.

    The year 2020 captures general discussions and political discussions relevant to the study.

    “We limit the data period to 2020 as it captures the general discussions, such as the pandemic (often non-political) and the political discussions related to the 2020 U.S. presidential election.”
2 of 2

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

  • Party-specific CPI counts do not sum to the reported total

    Major

    The reported Republican-to-Democrat and Democrat-to-Republican counts sum to 628,646, which is 33,015 fewer than the reported 661,661 CPI replies.

    Reported 661661 · recomputed 628646

  • Reported CPI share does not match the final reply count

    Major

    The paper reports 661,661 CPI replies and 1,594,271 replies in the final dataset. That is 41.5% of replies, not 34%.

    Reported 34 · recomputed 41.502417

2 of 2

Cites

The publisher hasn’t deposited this paper’s reference list.

Research works printed but not linked: 1

  1. C ¸etinkaya, Y. M.; Lee, Y.; Külah, E.; Toroslu, İ. H.; Cowan, M. A.; and Davulcu, H. 2024. Toward a Programmable Hu- manizing Artificial Intelligence Through Scalable Stance- Directed Architecture . IEEE Internet Computing, 28(05): 20-27.

Cited by

Sorted by citations

Related

Authors
Çetinkaya, Y. M., Ghafouri, V., Suarez-Tangil, G., Such, J., & Elmas, T.
Affiliations
  • Middle East Technical University (TR)
  • IMDEA Networks (ES)
  • University of Edinburgh (GB)
ORCID
Venue
Proceedings of the International AAAI Conference on Web and Social Media
Publisher
Association for the Advancement of Artificial Intelligence (AAAI)
Published
2025-06-07
Volume
19
Pages
324-340
ISSN
  • 2162-3449 (ISSN-L)
  • 2334-0770
DOI
10.1609/icwsm.v19i1.35819
OpenAlex
W4411120955
Cited by
2
References
0
Citations by year
2025: 2
FWCI
13.78
Type
conference paper
Language
en
Open access
Locations
  • Proceedings of the International AAAI Conference on Web and Social Media (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University) (cc-by)
Funders
  • Agencia Estatal de Investigación (13039, AEI/10)
  • European Social Fund
Topics
  • Social Sciences › Communication › Social Media and Politics
  • Social Sciences › Sociology and Political Science › Misinformation and Its Impacts
  • Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
Keywords
echo chambers, political polarization, Twitter, online toxicity, political discourse, 2020 US presidential election, content moderation, user engagement

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-13.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-13.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review