CoChat
Sign Up
Cited by 42Climate Change Communication and PerceptionSocial Media and PoliticsSentiment Analysis and Opinion Mining

Affective Polarization in Online Climate Change Discourse on Twitter

Open access
Tyagi, Aman, Uyheng, Joshua, Carley, Kathleen M.·2020·2020 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM)
DOI 10.1109/asonam49781.2020.9381419
Open this paper in CoChat — free accountOpen PDF

Abstract

Online social media has become an important platform to organize around different socio-cultural and political topics. An extensive scholarship has discussed how people are divided into echo-chamber-like groups. However, there is a lack of work related to quantifying hostile communication or affective polarization between two competing groups. This paper proposes a systematic, network-based methodology for examining affective polarization in online conversations. Further, we apply our framework to 100 weeks of Twitter discourse about climate change. We find that deniers of climate change (Disbelievers) are more hostile towards people who believe (Believers) in the anthropogenic cause of climate change than vice versa. Moreover, Disbelievers use more words and hashtags related to natural disasters during more hostile weeks as compared to Believers. These findings bear implications for studying affective polarization in online discourse, especially concerning the subject of climate change. Lastly, we discuss our findings in the context of increasingly important climate change communication research.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a network-based pipeline combining aspect-level sentiment, Krackhardt E/I indices, and Earth Mover’s Distance to measure affective polarization in online conversations. Applied to 100 weeks of Twitter climate change discourse, it finds climate change Disbelievers are more hostile toward Believers than vice versa, and Disbelievers use more natural disaster-related words/hashtags during more hostile weeks.

Key findings

  • Climate change Disbelievers showed relatively consistent hostility toward Believers over the 100-week period, with their affective polarization time series rarely falling below 0; 20 weeks were classified as hostile for Disbelievers (`l_k > 0.140`) versus 12 for Believers (`l_k > 0.080`).
  • Believers were not generally hostile toward Disbelievers, fluctuating above and below 0, though some weeks featured exceptionally high hostility; the standard deviation of `l_k` was lower for Disbelievers than for Believers.
  • Believers used more natural disaster-related hashtags than Disbelievers overall, but during exceptionally hostile weeks Believers used fewer such hashtags while Disbelievers used more; a greater percentage of Disbelievers’ tweets mentioned natural disaster-related words compared with Believers.
  • Stance labeling classified 3.9M users as Believers and 3.1M as Disbelievers out of 7M unique users, with average manual-validation precision of 81.8% across 1,000 sampled users.

Methods

  • Data: real-time tweets collected via Twitter’s standard API using keywords “Climate Change,” “#ActOnClimate,” and “#ClimateChange” from August 26, 2017 to September 14, 2019, with two server-error pauses ignored; after de-duplication, 38M unique tweets/retweets from 7M unique users were aggregated into 100 user-week periods.
  • Stance labels: a weak-supervision co-training model using label propagation and text classification, seeded with `ClimateChangeIsReal` and `SavetheEarth` for Believers and `ClimateHoax` and `Qanon` for Disbelievers; 500 users per group were manually validated.
  • Affective polarization metrics: Netmapper aspect-level sentiment scores from -1 to +1 toward mentioned agents; positive and negative affective networks; Krackhardt E/I indices; polarization valence `P_k = P-_k - P+_k`; and polarization magnitude `l_k` using Earth Mover’s Distance on weighted edge distributions, with hostile weeks defined as `l_k > mean + 1 SD`.

Limitations

  • The paper states that social media discourse, including Twitter, is not necessarily representative of a global population, and that the dataset excludes interactions not containing the collection keywords.
  • The paper states that aspect-level sentiment is non-trivial, that Netmapper is used without efforts to improve aspect-level sentiment scores, and that the natural disaster word list may miss specific disaster names such as “Dorian” instead of “Hurricane Dorian.”
  • The paper reports 81.8% average precision for stance labels, implying some misclassification (evident from the reported validation result); it also notes future work could expand the binary Believer/Disbeliever system.

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

Integrity signals

Registry

Crossref · 10.1109/asonam49781.2020.9381419

Cites

53 of 53 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

9 of 26 cited sources checked · 0 supported · 1 flagged · accuracy not yet measured

  • Premisecitation

    Existing scholarship on climate change discourse has relied on manually annotated corpora and datasets of limited size, facing barriers to measuring the emotional component of climate change discussions in a generalizable fashion.

    “Relying consistently on manually annotated corpora and datasets of limited size, existing scholarship has faced barriers to measuring the emotional component of climate change discussions in a generalizable fashion [8], [17], [18].”
    Quote in the paperSubstantial claim3 of 3 sources checked
    • cites [8] source

      The Discussion (page unavailable) says that locating a representative corpus of social-media discourse challenges established textual-analysis methods. Section 3.2 (page unavailable) describes the authors’ own manual sentiment classification of a subset of tweets. Neither passage establishes that existing scholarship consistently relied on manually annotated, limited-size datasets or that this specifically prevented generalizable measurement of emotion. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
    • cites [17] source

      The Introduction (p. 1) and Discussion and Related Work (pp. 11–12) address polarization on Indian and Pakistani social media, not climate change discourse. Although the Discussion notes that some prior polarization methods require annotations, our copy does not address the claimed reliance on small, manually annotated climate-change datasets or barriers to measuring emotion in those discussions (searched: climate change, emotion, annotations, dataset size). (LLM, conf=0.76)

      Experimental — accuracy not yet measured
    • cites [18] source

      The Data and Method section describes qualitative coding of testimony from 71 U.S. congressional hearings, supporting this paper as an example of a manually coded, bounded dataset. The Network Analysis Technique section says finer-grained judgments would have had low reliability and validity. Neither section establishes that existing scholarship consistently used small datasets or that this prevented generalizable measurement of emotion; this study coded policy stances, not emotion. Page numbers were not supplied in our copy. (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Findingown data

    Climate change Disbelievers are more hostile toward Believers in the anthropogenic cause of climate change than Believers are toward Disbelievers.

    “We find that deniers of climate change (Disbelievers) are more hostile towards people who believe (Believers) in the anthropogenic cause of climate change than vice versa.”
    • Also stated in: § IV. RESULTS · hedged
      “Overall, our analysis found that climate change Disbelievers tended to exhibit high levels of hostility toward climate change Believers.”
    Quote in the paperSubstantial claim
  • Findingown data

    Disbelievers use more words and hashtags related to natural disasters during more hostile weeks as compared to Believers.

    “Moreover, Disbelievers use more words and hashtags related to natural disasters during more hostile weeks as compared to Believers.”
    Quote in the paperSubstantial claim
  • Findingown data

    The high level of hostility of climate change Disbelievers toward Believers was relatively consistent throughout the 100-week period, with the Disbelievers' time series only very rarely going below the threshold of 0.

    “This finding was relatively consistent throughout the 100-week period under observation, as the time series for climate change Disbelievers only very rarely goes below the threshold of 0, which indicates similarly valenced interactions toward in-group and out-group members.”
    Quote in the paperSubstantial claim
  • Findingown data

    Some weeks displayed exceptionally high levels of hostility toward climate change Believers, greater than one standard deviation from the mean.

    “Some weeks displayed exceptionally high levels of hostility toward climate change Believers, greater than one standard deviation from the mean.”
    Quote in the paperNot a substantial claim
  • May combine two claims. 7
  • The quote only partly states this claim. 1
  • Classified as probably not a substantial claim. 9

  • May combine two claims.

    Premisecitation

    Online climate change discussions are interactionally polarized, exhibiting echo chamber-like homophilic interactions between Believers and Disbelievers.

    “In other words, online discussions about climate change are interactionally polarized, implying the persistence of echo chambers between Believers and Disbelievers [17]- [19].”
    • Also stated in: § B. Climate change and polarization · boosted
      “Studies employing social network analysis have uncovered robust evidence that online climate change discussions tend to exhibit echo chamber-like homophilic interactions [14], [19], [35].”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [14] source

      The Conclusion (page unavailable) explicitly says that describing online debate as an echo chamber is inaccurate: debaters report discussing with people who hold different values almost as often as with like-minded people. The paper examines political debate generally, not climate-change discussions or interactions between Believers and Disbelievers. (LLM, conf=0.84)

      Checked by a reasoning model — may be wrong.

      Experimental — accuracy not yet measured
    • cites [17] source

      The Results and Analysis (p. 9) describe an echo chamber-like pattern: pro-India and pro-Pakistan Twitter users had little interaction across positions. The paper studies discussion of India–Pakistan tensions, however, not climate change or interactions between climate Believers and Disbelievers; it cannot establish that pattern for those groups. (LLM, conf=0.88)

      Experimental — accuracy not yet measured
    • cites [19] source

  • May combine two claims.

    Premisecitation

    In rare instances of intergroup communication in online climate change discussions, more negative frames tended to prevail, including dismissal of climate change information as hoaxes, derailment of conversations to heated issues of identity, and overall higher levels of sarcasm and incivility.

    “Qualitative analysis further showed that in rare instances of intergroup communication, more negative frames tended to prevail, featuring dismissal of climate change information as hoaxes, derailment of conversations to heated issues of identity, as well as overall higher levels of sarcasm and incivility [8], [17], [18].”
    Quote in the paperNot a substantial claim3 of 3 sources checked
    • cites [8] source

      The Discussion says exposure to alternative views was relatively rare, and the Results report more negative sentiment in cross-group interactions. The paper does not substantiate the claimed qualitative findings about hoaxes, identity-related derailment, sarcasm, or incivility. (LLM, conf=0.87)

      Experimental — accuracy not yet measured
    • cites [17] source

      Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable, 1 not checked. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured
    • cites [18] source

      Split into 2 atoms — 0 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable, 1 not checked. Expand the claim to see per-atom verdicts.

      Experimental — accuracy not yet measured

  • May combine two claims.

    Premisecitation

    The hashtags #ClimateChangeIsReal and #SavetheEarth are used mostly by Believers, and #ClimateHoax and #Qanon are used mostly by Disbelievers.

    “These hashtags have been shown to be used mostly by the respective groups [14].”
    Quote in the paperSubstantial claim1 of 1 sources checked
    • cites [14] source

      The empirical-analysis passages report survey responses and attitude change in online debates, not hashtag use by climate-change Believers or Disbelievers. Searches for the four named hashtags and for climate-change hashtag use found no passage addressing the claimed group-specific usage. (LLM, conf=0.76)

      Experimental — accuracy not yet measured

  • May combine two claims.

    Findingown data

    Out of the total 7M users, the stance labeling algorithm labeled 3.9M as Believers and 3.1M as Disbelievers.

    “Out of the total 7M users, the algorithm labels 3.9M as Believers and 3.1M as Disbelievers.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    Earth Mover's Distance is able to capture more nuanced differences in distributions than other techniques such as the KS-Test.

    “However, during our experiments, we found that EMD is able to capture more nuanced differences in distributions.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The number of hostile weeks (l_k > 0.140) for Disbelievers is 20 and for Believers (l_k > 0.080) is 12.

    “The number of such weeks for Disbelievers where l k > 0.140 is 20 and for Believers where l k > 0.080 is 12.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The population can be divided into two groups, climate change Believers and Disbelievers, with opposing stances.

  • May combine two claims.

    Findingown data

    The mean plus 1 standard deviation for the affective polarization metric is -0.091 and 0.080 for Believers and -0.117 and 0.106 for Disbelievers.

    “The dotted lines represent mean 1 standard deviation, which for Believers is -0.091 and 0.080 and disbelievers is -0.117 and 0.106.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The population can be divided into two groups, climate change Believers and Disbelievers, with opposing stances.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given

    The proposed framework combines machine learning and network science methods in a novel, scalable, and generalizable fashion for ready application in a variety of contentious issues.

    “As the succeeding sections demonstrate, our framework combines machine learning and network science methods in a novel, scalable, and generalizable fashion for ready application in a variety of contentious issues.”
    Quote in the paperNot a substantial claim

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

    Premisecitation

    Generalized climate change beliefs over time are increasing.

    “Longitudinal analysis in prior work suggests that generalized climate change beliefs over time are increasing [15], [16]”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [15] source
    • cites [16] source

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

    Premisecitation

    Political elites or opinion leaders who share moralized content behave in distinct ways depending on their ideological orientations.

    “Prior work illustrates that these findings translate robustly to the digital sphere -political elites or opinion leaders who share moralized content behave in distinct ways depending on their ideological orientations [4].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [4] source

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

    Premisecitation

    Social media discourse does not necessarily constitute a representative sample of a particular global population.

    “Although social media discourse does not necessarily constitute a representative sample of a particular global population [44], digital platforms like Twitter nonetheless constitute a vital space for public conversations about important issues like climate change.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [44]

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

  • Classified as probably not a substantial claim.

    Premisecitation

    Online groups that disagree on a given topic are also more likely to be hostile toward each other.

    “Social scientific research examines how these phenomena are interconnected across a variety of contexts, such that online groups that disagree on a given topic are also more likely to be hostile toward each other [10].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [10] source

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

    Study design and conductcitation

    Netmapper has been used with reasonable accuracy for multiple sentiment level tasks.

    “We use Netmapper which has been used with reasonable accuracy for multiple sentiment level tasks [45], [46].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [45]

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

    • cites [46]

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

  • Classified as probably not a substantial claim.

    Premisecitation

    Ideological asymmetries underpin conflict dynamics across a variety of social issues.

    “Related scholarship in political psychology underscores how ideological asymmetries underpin conflict dynamics across a variety of social issues [42].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [42]

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

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    There is a lack of work related to quantifying hostile communication or affective polarization between two competing groups.

    “However, there is a lack of work related to quantifying hostile communication or affective polarization between two competing groups.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Interpretationcitation

    Such challenges may persist in the form of further entrenchment of Disbelievers within interactional siloes and disengagement from intergroup communication altogether.

    “Such challenges may thus persist in the form of further entrenchment of Disbelievers within interactional siloes and disengagement from intergroup communication altogether [19].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [19] source
16 of 16

  • The paper relies on this without support.

    The aspect-level sentiment score directed toward a user's Twitter handle accurately reflects the affective valence of the interaction between the two users.

    “For the purposes of this work, each tweet by a certain agent i which mentions or replies to agent j is assigned an aspect-level sentiment score from -1 (very negative) to +1 (very positive) directed toward the concept "@[agent j's Twitter handle]" [39].”

  • The authors state this as an assumption.

    The population can be divided into two groups, climate change Believers and Disbelievers, with opposing stances.

    “researchers may wish to expand the binary system of climate change beliefs assumed here.”

  • The authors state this as an assumption.

    The list of natural disaster-related words adequately captures communication about natural disasters.

    “Next, we use natural disaster-related words as a proxy to determine how natural disasters play a role in hostility between the two groups.”

  • The paper relies on this without support.

    Digital platforms like Twitter constitute a vital space for public conversations about important issues like climate change.

    “digital platforms like Twitter nonetheless constitute a vital space for public conversations about important issues like climate change.”
4 of 4

Self-check

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

Each contradiction between two passages listed here was confirmed by an independent check.

  • Disbeliever hostile-week cutoff conflicts with the figure caption

    Major

    The results define exceptionally hostile weeks as those above the mean plus one standard deviation and use 0.140 for Disbelievers. The Figure 1 caption gives their upper mean-plus-one-standard-deviation line as 0.106, so the stated cutoff and the resulting count of weeks use a different criterion.

  • Reported standard-deviation comparison is reversed

    Major

    The Figure 1 limits are described as mean ± one standard deviation. They imply a standard deviation of 0.1115 for Disbelievers versus 0.0855 for Believers, making the Disbelievers' standard deviation higher, not lower.

  • Negative-edge weight is zero for negative sentiment

    Minor

    The negative interaction network is said to contain negative-valenced interactions, but its stated weight rule, min(0, -x), assigns zero when sentiment x is negative. Instead, it assigns nonzero negative weights when x is positive.

  • Positive-edge weight is zero for positive sentiment

    Minor

    The positive interaction network is said to contain positive-valenced interactions, but its stated weight rule, min(0, x), assigns zero to every positive sentiment x. It therefore cannot weight those positive interactions as described.

4 of 4

Cites

Flagged first, then by citations

Research works printed but not linked: 1

  1. carley, 0, ORA & Netmapper, International Conference on Social Computing Behavioral-Cultural Modeling and Prediction and Behavior Representation in Modeling and Simulation

Cited by

Sorted by citations

Related

Authors
Tyagi, A., Uyheng, J., & Carley, K. M.
Affiliations
Carnegie Mellon University (US)
ORCID
Venue
2020 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM)
Publisher
IEEE
Published
2020-12-07
Pages
443-447
DOI
10.1109/asonam49781.2020.9381419
OpenAlex
W3082836236
Semantic Scholar
53087af154810e972a77493dd8701daf57531c44
Cited by
41
References
49
Citations by year
2026: 2 · 2025: 6 · 2024: 11 · 2023: 9 · 2022: 6 · 2021: 6 · 2020: 1
FWCI
12.06
Type
conference paper
Language
en
Open access
Licence
Locations
  • IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM) (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
Knight Foundation (N000141812106,N000141812108)
Topics
  • Social Sciences › Sociology and Political Science › Climate Change Communication and Perception
  • Social Sciences › Communication › Social Media and Politics
  • Physical Sciences › Computer Science › Artificial Intelligence › Sentiment Analysis and Opinion Mining
Keywords
affective polarization, climate change discourse, Twitter, climate change denial, echo chambers, social networking sites, hashtags, natural disasters, climate change communication

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