{"id":"6207f842-ad2d-4ce7-8538-79196a14fffb","arxiv_id":"2509.00696","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A comment queue that holds emotionally volatile comments until thread anger drops is claimed to reduce anger spread by 15%, but the reduction is produced by the queue rule itself, not by measured human behavior.","lead":"An online comment moderation system is proposed that scores each comment's emotional impact in a threaded conversation and briefly holds back comments that would push anger or fear past a threshold. The paper reports large reductions in toxicity and anger, but these come from a simulation, not from observing real users change their behavior.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline toxicity/anger reductions come from node deletion in simulation, not from observed self-reflection; causal claim unsupported.","rationale":"The paper's central claim requires that removing/queueing influential nodes in a conversation graph is a valid proxy for users engaging in self-reflection and changing their comments. This assumption is load-bearing: without it, the 12%/10% and 15% numbers are just a delete-and-measure exercise, not evidence for the proposed DER intervention. The manuscript nowhere tests the psychological mechanism; §3.5 operationalizes the intervention as node deactivation, and §4.1 explicitly postpones user feedback to future work. A computational baseline (e.g., random node removal) would only test selection quality, not the reflection mechanism; only an experiment with human participants can establish whether a delay plus emotion-board feedback changes user behavior. The reader's weakest assumption identifies the same gap, so I agree with the REJECT verdict and see no reason to change it.","tokens_in":10298,"tokens_out":6458,"duration_ms":77082,"concrete_test":"Run a preregistered randomized experiment: N≥200 participants respond to 10 emotionally charged posts sampled from the same Twitter/Reddit domains. In the treatment arm, participants see the eImpact emotion board and are required to wait 47 seconds before submitting, with an option to revise; in the control arm they post immediately. Score final comments with Perspective API and the NRC-based emotion board. If treatment does not produce significantly lower toxicity/anger than control, the simulation's node-deactivation results cannot be attributed to self-reflection.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central quantitative claims (12% toxicity reduction, 15% anger-spread reduction) are derived entirely from the simulation in §3.5, where reducing hate speech is implemented as 'deactivating influential nodes or restricting responses once they reach the toxicity threshold.' No user is ever observed reflecting or revising a comment; the queue mechanism (§3.4) delays publication, but the 'revision' step is a design proposal, not an empirical event. Thus the reported reductions measure the consequence of deleting/withholding high-impact nodes from a conversation graph, not the effect of the emotion-regulation intervention that the paper argues is responsible. The 12%/10% discrepancy between abstract and body further weakens precision. The paper itself defers 'real-time emotional feedback through user interviews and surveys' to future work (§4.1), acknowledging that the proposed psychological mechanism has not been tested with users. Because the independent variable (self-reflection) is never manipulated or measured, the evidence cannot support the causal claim that the framework reduces toxicity in real conversations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a graph-based framework (eImpact) that detects emotions in Twitter/Reddit conversations and uses a comment-queuing mechanism to delay comments that would push conversational anger/fear above thresholds. The intended psychological mechanism is that the delay induces self-reflection and emotion regulation, reducing toxicity. The authors report that eImpact reduces hate speech by 10% versus 7% for Google's Perspective API, that the queuing mechanism reduces the spread of anger/fear by 15%, and that only 4% of comments are held for about 47 seconds on average.","tokens_in":10562,"tokens_out":2857,"duration_ms":36140,"significance":"If the quantitative claims were valid, the work would address an important gap in online moderation: moving from post-hoc content removal to real-time, context-sensitive emotion regulation. The framework is creative, uses real social media data, and the comparison to Perspective API is a reasonable starting point. However, the evidence presented is not a measurement of human emotion regulation: it is a graph manipulation simulation in which nodes are deactivated or queued. The reported reductions are partly arithmetic consequences of the queuing rule. The inconsistency between the abstract's 12% and the body's 10% further erodes confidence. The potential significance is real, but the current manuscript does not establish it.","major_comments":[{"comment":"The abstract and introduction state a 12% toxicity reduction (e.g., 'reduced toxicity by 12%', '12% reduction in the spread of hate speech and anger'), while Section 3.5 and Table 1 report that the eImpact framework achieved a 10% reduction, with Perspective API at 7%. The paper does not explain which number is the measured result, or why the abstract differs from the body. This discrepancy directly affects the headline quantitative claim and must be resolved before the manuscript can be considered reliable.","section":"Abstract vs. Section 3.5 and Table 1"},{"comment":"The central causal claim — that promoting self-reflection reduces toxicity — is not tested. The experimental evaluation in Section 3.5 is described as 'deactivating influential nodes or restricting responses once they reach the toxicity threshold.' No user ever reflected on, revised, or even experienced the queue; the word 'revision' appears only as a design proposal in Section 3.4, and Section 4.1 explicitly defers 'real-time emotional feedback through user interviews and surveys' to future work. Thus the reported reductions measure the effect of node removal/queuing in a graph, not the effect of the emotion-regulation intervention that the paper argues is responsible.","section":"Section 3.5 and Section 4.1"},{"comment":"The reported 15% reduction in the spread of anger and fear is largely an arithmetic consequence of the queue rule: Section 3.4 states that comments are flagged and queued when they push Anger > 50% or Fear > 60%, and the results in Section 4 are computed on the resulting graph. Since queued comments cannot contribute to the emotion board until thresholds drop, measuring 'spread of anger and fear' on the output of the same rule does not provide independent evidence for emotion regulation. The manuscript also provides no error bars, statistical tests, or sensitivity analyses for the 15%, 10%, 4%, or 47-second figures, so the precision of these claims is unclear.","section":"Section 3.4 and Section 4"},{"comment":"The adaptive threshold mechanism is not specified precisely enough to be reproduced. The text says thresholds are 'determined using several parameters' and 'adjusted dynamically through an algorithm' but does not give the algorithm, the weighting of the parameters, or the exact rule for the sliding window (beyond '100 recent comments'), the weighted allowances for positive emotions, or the active/non-active distinction. Since the central results depend on these thresholds, the lack of specification is a load-bearing reproducibility gap.","section":"Section 3.4"}],"minor_comments":[{"comment":"The acknowledgement line 'We are grateful to Amity University for accepting our work as a keynote paper' is unclear and appears out of place. It may confuse readers about the venue of this submission.","section":"Section 5"},{"comment":"Reference [45] is the authors' prior conference paper and appears to be the basis for the framework and much of the experimental setup. The manuscript should clearly state what is new in this submission relative to [45], especially in the results and discussion.","section":"References"},{"comment":"The figures are described in the text but lack captions that specify the exact experimental condition, number of conversations, and whether values are means over conversations. Figure 5's histogram would benefit from error bars or confidence bands, and Figure 7's y-axis ('cumulative emotional impact') is never formally defined.","section":"Figures 5--7"},{"comment":"The emotion classification method is described as using the NRC lexicon and Emojinal embeddings, but the details of how text and emojis are combined into a single 0.1--1.0 intensity score are not given. A brief formula or algorithmic description would improve readability.","section":"Section 3.2"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be a light extension of the authors' prior IEEE Engineering Informatics paper (ref [45]), reusing its framework and evaluation, with the queuing mechanism as an addition. The core quantitative claims are not empirically grounded: the simulation deletes nodes rather than studying human reflection, and the 15% anger reduction is definitionally tied to the queue rule. Unless the authors reframe the contribution as a purely algorithmic simulation study and substantially weaken the causal language — which would be a major change in scope — the paper is unlikely to meet the standards of the journal. I would recommend rejection, though the underlying idea could be developed in a future study with actual user interaction and behavioral measures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The queue concept is plausible and the paper describes it clearly enough to implement. The useful bit is the adaptive piece: holding comments that would push anger/fear past thresholds, then re-evaluating them as new comments arrive, with context-sensitive threshold adjustment. That is a real extension of the earlier eImpact work and of fixed-delay ideas. If someone built this into a platform and measured what happens, that would be a meaningful contribution.\n\nThe problems are in the evidence, not the idea. The abstract says 12% toxicity reduction; the body says 10% (Section 3.5). More importantly, the 15% anger/fear reduction and the 10% toxicity reduction come from running the proposed rule itself: comments that exceed thresholds are held, so they cannot influence the emotion board until thresholds drop. Measuring the spread of anger on the output of that same rule is close to a tautology. No user ever reflected, revised, or changed behavior; the paper itself says real-time emotional feedback through user interviews is future work (Section 4.1). The weakest assumption is that deactivating an influential node in a conversation graph is a model of human self-reflection. That assumption is doing all the work.\n\nThere are also no error bars, no sensitivity analyses, and no code or data to check how much the free parameters (thresholds, window size, influence weights) drive the results. The held-duration histograms and emotion distributions are descriptive, not inferential.\n\nI do not think the authors are being deceptive—the gap between the mechanism and the psychological claim is the kind of thing that happens when computer scientists simulate an intervention and then describe it in causal language. But as it stands, the central quantitative claims are not supported as measurements of human behavior.\n\nI would send this to peer review rather than desk reject. The design is fresh and testable, and a good referee report could push the authors to run the actual user study or at least reframe the results as simulation outcomes. I would not cite the numbers in my own work, and I would not bring it to reading group as a model of evaluation. But as a design direction, it is worth engaging with seriously.","headline":"A clear, testable queue design is buried under empirical claims the paper's own simulation cannot support; worth a referee, not a citation.","tokens_in":11011,"tokens_out":1750,"would_cite":false,"duration_ms":26163,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that combining a graph-based emotion board with a comment queue can reduce online toxicity by 12% and anger spread by 15%, while holding only 4% of comments.","keywords":["digital emotion regulation","online toxicity","self-reflection","comment queuing","conversation graph","emotion contagion","content moderation","Twitter and Reddit"],"falsifier":"Run a randomized controlled field test on a live comment platform: for comments flagged by the emotion board, randomly assign authors to immediate publication, queue with a reflection prompt, or queue without a prompt. If the prompted-and-queued group's subsequent comments show no larger drop in toxicity or anger than the immediate group, the central claim that self-reflection and delay reduce toxicity is falsified. A complementary simulation check: recompute the reported reductions while randomly holding the same 4% of comments instead of targeting high-influence ones; if random holds achieve","tokens_in":10206,"feed_emoji":"⏳","tokens_out":6368,"duration_ms":72944,"temperature":0.7,"pith_summary":"This paper tries to establish that online toxicity can be reduced in real time by giving users a chance to reflect, rather than only deleting harmful comments after the fact. It proposes a graph-based emotion-regulation framework that scores each comment's influence on the emotional tone of a conversation, plus a comment queue that briefly holds comments likely to push anger or fear past dynamic thresholds. Analyzing Twitter and Reddit conversations, the paper reports that the framework reduced toxicity by 12% (10% in the body's comparison, versus 7% for the Perspective API), that the queue reduced anger and fear spread by 15%, and that only 4% of comments were held, for an average of 47 seconds. If these numbers hold in deployed systems, platforms could meaningfully lower user-perceived hostility without substantially slowing conversation.","feed_headline":"Graph plus comment queue cuts toxicity by up to 12%","feed_subtitle":"Self-reflection prompts and brief holds beat Perspective API's 7% while delaying only 4% of posts.","key_machinery":"The load-bearing object is the conversation graph with its emotion board at the root, plus the adaptive comment queue. The graph assigns each comment an emotion and intensity from the NRC lexicon and computes influence from reply count, distance to root, PageRank, and emotion intensity; the emotion board aggregates these into percentages. The queue holds comments that push anger or fear past dynamic thresholds, re-evaluates them as new comments arrive, and either releases them when emotional balance is restored or prompts revision or suspension. Dynamic thresholds, a sliding window over recent comments, and weighted allowances for positive and underrepresented emotions keep the queue from st","core_discovery":"The central claim is that emotion regulation can be operationalized as a graph-computational problem. Each conversation is modeled as a directed acyclic graph whose root is the original post; comments are nodes scored for emotion and intensity using the NRC lexicon, and each node's influence is computed from number of replies, distance from root, PageRank, and emotional intensity. The root's emotion board tracks the percentage contribution of each emotion. A comment whose addition would push anger above 50% or fear above 60% is temporarily queued; it is re-evaluated after each new comment and can be reintegrated when subsequent positive or balancing comments reduce the threat. The paper repo","pith_inferences":["I would expect the reported 12% and 15% effects to shrink or shift in a live deployment, because the simulation assumes a held or removed node is equivalent to a human author deciding to self-reflect; a test with real authors editing after a prompt is the natural next step.","An ablation that queues only the highest-influence nodes would reveal whether the benefit comes from a few central comments or from the breadth of holds; that result could justify lower hold rates than 4%.","The graph influence score could be turned into a pre-publication preview, such as 'this comment will raise anger by X percent,' which might reduce the need for actual holds—an application the paper gestures at but does not test.","Because the emotion board reacts to contextual balance, the framework could be repurposed to amplify underrepresented emotions like joy and trust in polarized conversations, not just suppress anger."],"forward_implications":["Conversation context matters for moderation: scoring a comment by its position, replies, and PageRank can identify toxicity that text-only detectors miss.","A short, adaptive delay can reduce anger propagation with little conversational cost, since only about 4% of comments are held.","Dynamic thresholds let moderation intensity track conversation heat, so active threads do not stall and calm threads stay strict.","Combining self-reflection prompts with delayed publication gives platforms a proactive alternative to post-hoc deletion.","Emotion-targeted holding of anger and fear can shift a conversation's emotional balance toward positive emotions, as the emotion-board comparisons show."],"supporting_citations":[{"why":"Supplies the NRC emotion lexicon used to assign each comment its emotion and intensity.","marker":"[32]"},{"why":"Provides the Perspective API toxicity scores used as the baseline the framework must beat.","marker":"[11]"},{"why":"Earlier eImpact framework whose graph-influence scoring and conversation representation this paper extends.","marker":"[45]"},{"why":"Shows temporal distancing reduces emotional arousal, the psychological basis for the queuing delay.","marker":"[17]"},{"why":"Motivates self-reflection and emotion-regulation design in online news commenting, the intervention's conceptual source.","marker":"[23]"},{"why":"Establishes affect labeling as implicit emotion regulation, supporting the emotion-board feedback mechanism.","marker":"[43]"},{"why":"Documents trolling as driven by mood and context, justifying a delay and reflection mechanism rather than pure removal.","marker":"[5]"},{"why":"Defines emotion regulation as managing experience, expression, and timing, the theoretical frame for the framework.","marker":"[14]"},{"why":"Defines Digital Emotion Regulation and situates the work within human-computer interaction.","marker":"[46]"}],"fun_headline_variants":["Queuing comments reduces anger spread by 15%","Graph-based emotion check cuts online toxicity 12%","Brief hold on comments lowers anger and toxicity","Emotion-aware queue trims toxic posts by 12%","Delayed comments cut anger spread 15%, toxicity 12%"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The reported reductions assume that deactivating influential nodes in a conversation graph is equivalent to a real user noticing their emotional impact and choosing to self-reflect or revise; no participant in the study actually made that choice.","fun_headline_variants_meta":{"raw":{"variants":["Queuing comments reduces anger spread by 15%","Graph-based emotion check cuts online toxicity 12%","Brief hold on comments lowers anger and toxicity","Emotion-aware queue trims toxic posts by 12%","Delayed comments cut anger spread 15%, toxicity 12%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000194,"raw_usage":{"total_tokens":1170,"prompt_tokens":706,"completion_tokens":464,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":385}},"tokens_in":450,"tokens_out":464,"duration_ms":6035,"temperature":1.0,"reasoning_tokens":385,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T13:18:43.593204+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a randomized controlled field test on a live comment platform: for comments flagged by the emotion board, randomly assign authors to immediate publication, queue with a reflection prompt, or queue without a prompt. If the prompted-and-queued group's subsequent comments show no larger drop in toxicity or anger than the immediate group, the central claim that self-reflection and delay reduce toxicity is falsified. A complementary simulation check: recompute the reported reductions while randomly holding the same 4% of comments instead of targeting high-influence ones; if random holds achieve","supporting_citations":[{"cited_title":"National Research Council, Canada 2, 234 (2013)","cited_arxiv_id":null,"evidence_quote":"Supplies the NRC emotion lexicon used to assign each comment its emotion and intensity."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Perspective API toxicity scores used as the baseline the framework must beat."},{"cited_title":"In: 2023 IEEE Engineer- ing Informatics","cited_arxiv_id":null,"evidence_quote":"Earlier eImpact framework whose graph-influence scoring and conversation representation this paper extends."},{"cited_title":"Psychological science25(8), 1571–1580 (2014)","cited_arxiv_id":null,"evidence_quote":"Shows temporal distancing reduces emotional arousal, the psychological basis for the queuing delay."},{"cited_title":"In: Proceedings of the 25th Interna- tional Academic Mindtrek Conference","cited_arxiv_id":null,"evidence_quote":"Motivates self-reflection and emotion-regulation design in online news commenting, the intervention's conceptual source."},{"cited_title":"Emotion Review10(2), 116–124 (2018)","cited_arxiv_id":null,"evidence_quote":"Establishes affect labeling as implicit emotion regulation, supporting the emotion-board feedback mechanism."},{"cited_title":"In: Proceedings of the 2017 ACM conference on computer supported cooperative work and social computing","cited_arxiv_id":null,"evidence_quote":"Documents trolling as driven by mood and context, justifying a delay and reflection mechanism rather than pure removal."},{"cited_title":"Handbook of emotions3(3), 497–513 (2008)","cited_arxiv_id":null,"evidence_quote":"Defines emotion regulation as managing experience, expression, and timing, the theoretical frame for the framework."},{"cited_title":"Current Directions in Psychological Science29(4), 412–418 (2020)","cited_arxiv_id":null,"evidence_quote":"Defines Digital Emotion Regulation and situates the work within human-computer interaction."}],"review_version":1}