Pith. sign in

REVIEW 3 major objections 6 minor 22 references

SPILLOVER: Measuring Cyberbullying NormPropagation on Social Media

T0 review · 3 major / 6 minor · reviewed 2026-08-01 · deepseek-v4-flash

Pith's one-line read This paper claims that cyberbullying transmits from comment to comment within social media sessions, and that the transmission is content-specific rather than generic disinhibition.

desk verdict Solid comment-pair study of cyberbullying spillover with a robust core result; the cross-user norm-propagation claim needs a cross-user-only odds ratio before it fully lands. read the letter →

arxiv 2607.19646 v1 pith:YQKA6LQP submitted 2026-07-22 cs.SI

classification cs.SI
keywords cyberbullyingspillovernormpropagationsocialmediacommentpairscontentsimilarityemotionalcontagionmoraldisengagement
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

This paper tries to establish that cyberbullying is not just a property of individual comments but a contagious process that passes from one comment to the next within a conversation. Using tens of thousands of consecutive comment pairs from Instagram and three other platforms, it finds that a preceding cyberbullying comment roughly doubles the odds the next comment will also be cyberbullying, even after accounting for how toxic the whole session is. The paper also claims the transfer is content-specific: not merely more aggression, but the same kind of aggression, with similar vocabulary and style. A single binary feature—whether the previous comment was cyberbullying—improves automated detection beyond text-only classifiers, suggesting a practical moderation signal. The authors are careful to note that observational data cannot fully rule out an unobserved common trigger within a session.

What carries the argument

The unit of analysis is the consecutive comment pair (c_{i-1}, c_i), ordered by timestamp within a session, classified by whether each comment is cyberbullying. The argument leans on three controls: a session fixed-effects logistic regression that makes each session its own control, a stratified analysis by session toxicity, and a matched cross-session baseline that pairs each within-session cyberbullying-to-cyberbullying pair with a same-toxicity, same-length cross-session pair to rule out shared vocabulary and shared topic. Five text similarity methods (TF-IDF cosine, word Jaccard, character 3-gram, and two sentence encoders) provide convergent evidence for content transfer.

What would settle it

A controlled experiment that randomly assigns users to see either a cyberbullying comment or a neutral comment in otherwise identical threads and then measures the toxicity of the next comment would settle it: if the next-comment aggression rate does not differ between conditions, the spillover claim fails. Alternatively, showing that a within-session common cause (e.g., a sudden influx of new users) produces the same dose-response and similarity patterns would undermine the causal reading.

Watch

Extended reading notes

Core claim

The central claim is that cyberbullying spills over immediately from comment to comment: within the 430 Instagram sessions that contain at least one cyberbullying comment, a comment preceded by a cyberbullying comment is cyberbullying 49.3% of the time versus 20.5% after a non-cyberbullying comment, a 28.8 percentage-point gap that survives session fixed effects (odds ratio 1.83). The spillover is mostly cross-user (82.9% of consecutive cyberbullying comments come from different users) and content-specific, as cyberbullying-to-cyberbullying pairs are substantially more textually similar than non-cyberbullying-to-cyberbullying pairs across five lexical and semantic measures, and this similari

Load-bearing premise

The central premise is that within a session there is no unobserved time-varying trigger—like a viral post or a trending controversy—that independently makes consecutive comments aggressive and textually similar, because such a common cause would mimic spillover.

Editorial extensions

If this is right

  • If the spillover is causal, moderation can target the spreading chain rather than only individual offenders; a single prior-comment flag improves average precision over session-level baselines and over a fine-tuned text classifier.
  • Because the transfer is content-specific, exposure to one type of attack (e.g., appearance-based) raises the chance of the same type of attack next, not just aggression in general.
  • Platform design matters: linear comment layouts show stronger spillover (up to 1.95x) than threaded layouts (1.26x), implying visibility drives transmission.
  • Even non-aggressive replies turn more negative after severe cyberbullying, so the effect operates below the threshold of overt aggression.
  • The spillover signal persists two to three steps beyond the triggering comment, though it decays.

Reading between the lines

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

  • If spillover is causal, then a randomized intervention that delays or relabels a cyberbullying comment before it is widely seen should measurably reduce the aggression rate of subsequent comments; the paper itself calls for such an experiment.
  • The topic heterogeneity (appearance-based attacks showing the highest spillover, odds ratio about 2.1) suggests targeted moderation that watches for high-spillover topic categories after an initial attack.
  • The finding that a single binary predecessor feature is orthogonal to comment text implies that content-based classifiers under-use conversational context; combining both dimensions could yield larger gains in production systems.
  • Adversarial adaptation is a foreseeable limitation: if users learn that inserting a neutral comment resets the detection signal, they could break the chain; studying such gaming behavior would clarify the signal's real-world durability.
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

3 major / 6 minor

Summary. The paper analyzes comment-to-comment cyberbullying (CB) spillover using 32,754 consecutive comment pairs from 430 Instagram sessions that contain at least one CB comment. The authors report that a preceding CB comment raises the probability that the next comment is CB by 28.8 percentage points, with a session fixed-effects odds ratio of OR=1.83 (p<10^-61), and they claim this effect is driven primarily by cross-user spread (82.9% of CB→CB pairs are from different users). They further find that CB→CB pairs are more textually similar than NoCB→CB pairs across five similarity methods, that the effect survives a matched cross-session baseline, that non-CB replies grow more negative with preceding CB severity, and that a single binary feature (prior comment is CB) improves prediction over session-level baselines and a fine-tuned HateBERT classifier. They replicate the occurrence and similarity findings on Reddit, Wikipedia Detox, and SOCC using automated toxicity labels.

Significance. If the central claims hold, the paper makes a useful contribution: it moves beyond documenting clustering of cyberbullying to measuring immediate comment-to-comment transmission, with content-specific copying and a practical moderation signal. The strengths of the paper include the multi-method similarity analysis (five methods, including two neural encoders), the matched cross-session baseline, the dose-response sentiment analysis, the cross-platform replication, and the availability of code and data. The occurrence-level association is robust to session fixed effects and permutation tests, and the prediction experiments are carefully framed with AUPRC and held-out folds. However, the headline "driven primarily by cross-user spread" claim is not yet supported by the reported analyses, and the causal interpretation remains vulnerable to a within-session time-varying confound that the authors acknowledge. These issues are load-bearing for the paper's distinguishing contribution, so the appropriate revision is major.

major comments (3)
  1. [§3.2, Table 2] The abstract and §3.2 claim that the spillover effect is "driven primarily by cross-user spread." This is not established by the evidence presented. Table 2 shows that 82.9% of CB→CB pairs are cross-user by volume, but the per-pair CB rate is higher for same-user pairs (59.9%) than for cross-user pairs (31.2%). The fixed-effects OR (Eq. 2, OR=1.83) is computed on all pairs, so same-user persistence could inflate the aggregate effect. Please report a session fixed-effects OR restricted to different-user pairs, and ideally an interaction between prior CB status and same/different user. The role-label analysis used to argue that second commenters are "new aggressors" also relies on labels (bully, victim) that are derived from the same CB behavior annotations, making the inference circular. An independent measure of the second commenter's behavior, or a label not defined by the outcome, is n
  2. [§4, Limitations (within-session time-varying confound)] The authors correctly acknowledge that a viral post or trending controversy could drive many users to comment aggressively in rapid succession without any transmission between comments. This is not a minor caveat. The session fixed-effects model (Eq. 2) absorbs between-session heterogeneity but not within-session shocks, so the OR=1.83 could reflect a common trigger rather than comment-to-comment influence. The dose-response gradient (Fig. 1) and the matched similarity baseline are suggestive, but they address sentiment and text similarity, not the occurrence-level transition itself. Please add within-session analyses that can distinguish a shared trigger from transmission: for example, include comment position or time-window fixed effects, compare against sessions with no CB but similar engagement spikes, or run a placebo test using the next-following comment as a "previous" comment (re
  3. [§3.6, Cross-Platform Replication] The three external datasets are labeled with Detoxify at a 0.5 threshold. Detoxify measures generic toxicity, not cyberbullying, which requires targeting, repetition, and power imbalance. The authors note this caveat in §3.6, but Table 6 and the abstract's "key findings replicate across three independent datasets" present these as cyberbullying spillover. Please relabel the external results as toxicity spillover, or provide evidence that Detoxify-positive comments in these datasets satisfy the cyberbullying definition beyond AUC=0.91 against human severity labels, which does not address target or repetition. The threshold robustness checks are a good start but do not fix the construct validity issue.
minor comments (6)
  1. [Table 3 caption] The caption says "All p<10^-100" without specifying the test or the multiple-comparison correction. Also, "XUd" is used for cross-user pairs but the abbreviation is not defined.
  2. [§3.1] The analysis is restricted to 430 sessions containing at least one CB comment. Please clarify that the base rates in Table 1 are conditional on this restriction and are not representative of the full Instagram corpus.
  3. [§3.2, Eq. (2)] The session fixed-effects logistic regression with one intercept per session may suffer from incidental-parameter bias when session sizes are moderate. Consider reporting a conditional logistic regression or a robustness check with cluster-robust standard errors.
  4. [Fig. 1] The figure reports d values for severity levels 1, 2, and 3 relative to severity 0. The caption or text should note the reference category and that the Spearman rho=0.112 is computed across all comments.
  5. [§4, Discussion] The text states that "the spillover effect does not decay with the time gap between comments," but no analysis or table is provided for this claim. Please add the supporting result or remove the statement.
  6. [§3.1, role labels] The seven role labels (bully, victim, defender, bystander, etc.) are used in the cross-user analysis, but no definition, annotation protocol, or inter-annotator agreement is reported. Please provide these details.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the development is self-contained with held-out prediction and resampling controls.

full rationale

The paper's central claims are not derived from their own inputs by construction. The occurrence-level spillover (Eq. 1, Table 1) is a raw conditional comparison, and the within-session odds ratio (Eq. 2, OR=1.83) is estimated with session fixed effects; neither is a fitted parameter renamed as a prediction. The content-similarity results (Table 3) are explicitly controlled by a matched cross-session baseline that replaces the following CB comment with a CB comment from a different session, directly removing the shared-vocabulary and session-composition confounds. The prediction experiments use 5-fold session-grouped cross-validation, and the prev_is_cb feature is evaluated on held-out folds and as an increment over session baselines and a fine-tuned HateBERT model; this is legitimate feature evaluation, not circular fitting. Cross-platform replication uses external datasets with automated labels validated against human annotations (AUC=0.91), so it does not reduce to the paper's own claims. The authors' self-citations ([4], [5]) appear only in the related-work survey and are not load-bearing for any main result. The role-label interpretation that cross-user second commenters are mostly 'bullies or bully assistants' is a validity caveat—role labels may not be fully independent of CB status—but the paper does not define those roles as an algebraic transform of the CB labels, so this is not a demonstrated circular reduction. Overall, the derivation chain is self-contained; the stated within-session time-varying confound is acknowledged as a limitation rather than hidden by the analysis.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claims rest on a small set of hand-chosen thresholds (CB severity threshold, Detoxify threshold) and on the assumption that no within-session time-varying trigger drives both consecutive comments. No new theoretical entities are introduced.

free parameters (4)
  • CB label threshold (severity >= 1) = 1 on 0-3 scale
    A comment is classified as CB if the majority of five annotators assign severity >=1. This hand-chosen threshold defines the outcome and directly determines all spillover rates; no robustness analysis varying this threshold is reported.
  • Detoxify toxicity threshold = 0.5 (robustness at 0.3 and 0.7)
    Applied to Reddit, Wikipedia Detox, and SOCC to generate proxy labels; threshold choice affects which comments count as toxic, though the paper reports ORs within 0.08 across thresholds.
  • Session inclusion criterion = >=1 CB comment (430 of 1,559 sessions)
    All analyses are scoped to sessions where at least one CB comment exists; this selection is intentional and disclosed, but all claims are conditional on it.
  • Minimum session length for fixed-effects model = 20 comments (robustness at 10)
    The fixed-effects analysis is restricted to these sessions; the number of sessions in the subset is not reported, and the OR may differ in shorter sessions.
assumptions (5)
  • domain assumption Human severity annotations (majority of five annotators) are a valid ground truth for cyberbullying status
    All Instagram analyses treat these labels as correct; no inter-annotator agreement or label-quality analysis is reported.
  • domain assumption No unobserved within-session time-varying confounder drives consecutive CB comments and their similarity
    The causal reading of the dose-response and similarity results requires this; the authors explicitly note the 'viral post' confound in Limitations and cannot rule it out.
  • domain assumption Timestamp order in flat Instagram comments reflects conversational influence
    Pairs are consecutive by timestamp, but flat comment sections do not encode reply structure; a CB comment and a later comment may both independently respond to the post.
  • domain assumption Text similarity operationalizes content transfer
    Elevated similarity between CB→CB pairs is taken as evidence of copying; shared target or formulaic abuse are alternative drivers.
  • domain assumption Detoxify toxicity labels are a valid proxy for cyberbullying on external platforms
    Cross-platform replication uses Detoxify at 0.5; validated on Instagram with AUC=0.91, but toxicity is broader than cyberbullying.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SPILLOVER: Measuring Cyberbullying NormPropagation on Social Media." pith.science (2026). https://pith.science/paper/YQKA6LQP

@misc{pith2026260719646,
  author       = {Pith},
  title        = {Pith review of: SPILLOVER: Measuring Cyberbullying NormPropagation on Social Media},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YQKA6LQP}},
  note         = {Machine review of arXiv:2607.19646}
}
abstract

While certain aspects of cyberbullying (CB) such as its factors and prevalence have been studied extensively, relatively little attention has been given to specifically how the aggression transfers from comment to comment. This understanding could have important implications for designing better anti-bullying features. In this paper, we study multiple aspects of the nature of this aggression transference in social media sessions. Using data from 32,754 consecutive comment pairs from 430 Instagram sessions, we find that a preceding CB comment substantially raises the odds of the next comment being CB, an effect confirmed by session fixed-effects controls and driven primarily by cross-user spread. We also find that $\text{CB} \to \text{CB}$ pairs are more textually similar than $\text{NoCB} \to \text{CB}$ pairs across five complementary methods, and that this pattern holds under a matched cross-session baseline that rules out shared vocabulary, session toxicity, and session length as confounds. Moreover, non-aggressive replies grow more negative as preceding CB severity increases, a graded pattern consistent with automatic emotional influence below the threshold of overt aggression. These key findings replicate across three independent datasets (Reddit, Wikipedia Detox, and SOCC), with spillover rates that track platform visibility design. Finally, we show that a single binary feature (whether the prior comment was CB) improves prediction over session-level baselines and over a fine-tuned HateBERT classifier, serving as a real-time moderation signal that targets the spreading chain rather than individual offenders.

Figures

Figures reproduced from arXiv: 2607.19646 by the authors.

Figure 1
Figure 1. Dose-response gradient: mean sentiment of NoCB replies as a function of pre￾ceding comment severity. Sentiment grows monotonically more negative; the effect size between severity 0 and severity 3 is d = 0.345. 3.3 Sentiment Shift and Dose-Response If the spreading is a genuine contagion process rather than generic disinhibition, then exposure to CB should affect the emotional tone of subsequent comments, even those … view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 5 canonical work pages

  1. [1]

    In: Advances in Information Retrieval: 40th European Conference on IR Research, ECIR 2018

    Agrawal, S., Awekar, A.: Deep learning for detecting cyberbullying across multiple social media platforms. In: Advances in Information Retrieval: 40th European Conference on IR Research, ECIR 2018. pp. 141–153 (2018)

  2. [2]

    Per- sonality and Social Psychology Review3(3), 193–209 (1999)

    Bandura, A.: Moral disengagement in the perpetration of inhumanities. Per- sonality and Social Psychology Review3(3), 193–209 (1999)

  3. [3]

    In: Proceedings of the International AAAI Con- ference on Web and Social Media

    Baumgartner, J., Zannettou, S., Keegan, B., Squire, M., Blackburn, J.: The pushshift Reddit dataset. In: Proceedings of the International AAAI Con- ference on Web and Social Media. vol. 14, pp. 830–839 (2020)

  4. [4]

    In: Proceedings of the International AAAI Con- ference on Web and Social Media (ICWSM)

    Bisharat, A., Sandoval, M., Nazari, M., Hall, D.L., Abuhamad, M., Silva, Y.N.: SpectrumNet: Detecting LGBTQ+ cyberbullying with dynamic context-aware attention. In: Proceedings of the International AAAI Con- ference on Web and Social Media (ICWSM). vol. 20, pp. 276–290 (2026). https://doi.org/10.1609/icwsm.v20i1.42638

  5. [5]

    arXiv preprint arXiv:2409.12263 (2024)

    Bisharat, A., Sandoval Madrigal, M., Abuhamad, M., Hall, D.L., Silva, Y.N.: Detecting LGBTQ+ instances of cyberbullying. arXiv preprint arXiv:2409.12263 (2024)

  6. [6]

    Journal of School Violence14(1), 30–46 (2015)

    Bussey, K., Fitzpatrick, S., Raman, A.: The role of moral disengagement and self-efficacy in cyberbullying. Journal of School Violence14(1), 30–46 (2015)

  7. [7]

    In: Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021)

    Caselli, T., Basile, V., Mitrović, J., Granitzer, M.: Hatebert: Retraining bert for abusive language detection in english. In: Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021). pp. 17–25 (2021)

  8. [8]

    In: Pro- ceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing

    Cheng, J., Bernstein, M., Danescu-Niculescu-Mizil, C., Leskovec, J.: Anyone can become a troll: Causes of trolling behavior in online discussions. In: Pro- ceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing. pp. 1217–1230 (2017)

Show all 22 references
  1. [9]

    In: Proceedings of the International AAAI Conference on Web and Social Media

    Dinakar, K., Reichart, R., Lieberman, H.: Modeling the detection of textual cyberbullying. In: Proceedings of the International AAAI Conference on Web and Social Media. vol. 5, pp. 11–17 (2011)

  2. [10]

    Hanu, L., Unitary team: Detoxify.https://github.com/unitaryai/ detoxify(2020)

  3. [11]

    Hatfield,E.,Cacioppo,J.T.,Rapson,R.L.:Emotionalcontagion.Cambridge University Press (1994)

  4. [12]

    arXiv preprint arXiv:1503.03909 (2015)

    Hosseinmardi, H., Mattson, S.A., Rafiq, R.I., Han, R., Lv, Q., Mishra, S.: Detection of cyberbullying incidents on the instagram social network. arXiv preprint arXiv:1503.03909 (2015)

  5. [13]

    Corpus Pragmatics4(2), 155–190 (2020) SPILLOVER: Measuring Cyberbullying Norm Propagation on Social Media 15

    Kolhatkar, V., Wu, H., Cavasso, L., Francis, E., Shukla, K., Taboada, M.: The SFU opinion and comments corpus: A large-scale dataset with anno- tations for stance, constructiveness and toxicity. Corpus Pragmatics4(2), 155–190 (2020) SPILLOVER: Measuring Cyberbullying Norm Prop...

  6. [14]

    Proceedings of the National Academy of Sciences111(24), 8788–8790 (2014)

    Kramer, A.D., Guillory, J.E., Hancock, J.T.: Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences111(24), 8788–8790 (2014)

  7. [15]

    Paciello, M., Tramontano, C., Nocentini, A., Fida, R., Menesini, E.: The role of traditional and online moral disengagement on cyberbullying: Do externalising problems make any difference? Computers in Human Behavior 103, 190–198 (2020).https://doi.org/10.1016/j.chb.2019.09.024

  8. [16]

    SSRN Preprint (2025).https://doi.org/10.2139/ssrn.5167003

    Prama, T.T., Amrin, J.F., Anwar, M.M., Sarker, I.H.: Ai-enabled user- specific cyberbullying severity detection with explainability. SSRN Preprint (2025).https://doi.org/10.2139/ssrn.5167003

  9. [17]

    In: Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing

    Reimers, N., Gurevych, I.: Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In: Proceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing. pp. 3982–3992 (2019)

  10. [18]

    Social Psychological & Personality Science 10(8), 1028–1035 (2019).https://doi.org/10.1177/1948550618820309

    Rosenbusch, H., Evans, A.M., Zeelenberg, M.: Multilevel emotion trans- fer on YouTube: Disentangling the effects of emotional contagion and ho- mophily on video audiences. Social Psychological & Personality Science 10(8), 1028–1035 (2019).https://doi.org/10.1177/1948550618820309

  11. [19]

    In: Advances in Neural Informa- tion Processing Systems

    Song, K., Tan, X., Qin, T., Lu, J., Liu, T.Y.: MPNet: Masked and permuted pre-training for language understanding. In: Advances in Neural Informa- tion Processing Systems. vol. 33, pp. 16857–16867 (2020)

  12. [20]

    PLoS ONE15(10), e0240924 (2020).https://doi.org/ 10.1371/journal.pone.0240924

    Talpur, B.A., O’Sullivan, D.: Cyberbullying severity detection: A machine learning approach. PLoS ONE15(10), e0240924 (2020).https://doi.org/ 10.1371/journal.pone.0240924

  13. [21]

    In: Proceedings of the 26th International Conference on World Wide Web

    Wulczyn, E., Thain, N., Dixon, L.: Ex machina: Personal attacks seen at scale. In: Proceedings of the 26th International Conference on World Wide Web. pp. 1391–1399 (2017).https://doi.org/10.1145/3038912.3052591

  14. [22]

    In: Proceedings of the 2012 Conference of the North American ChapteroftheAssociationforComputationalLinguistics:HumanLanguage Technologies

    Xu, J.M., Jun, K.S., Zhu, X., Bellmore, A.: Learning from bullying traces in social media. In: Proceedings of the 2012 Conference of the North American ChapteroftheAssociationforComputationalLinguistics:HumanLanguage Technologies. pp. 656–666 (2012)

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.