{"id":"dbfbee8b-b3eb-4a3b-83bb-aa49eabbed19","arxiv_id":"2506.00583","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-step LLM pipeline that identifies and replaces offensive or intensifying emojis in tweets reduces perceived offensiveness in human evaluation, especially for mild offenses, without large semantic loss.","lead":"This paper studies how emojis make Twitter posts more offensive, and proposes an AI pipeline that replaces harmful emojis with safer ones. The authors show the approach lowers perceived offensiveness in human tests, mainly for mildly offensive posts.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pipeline's emoji-only manipulation is unverified: the offensiveness drop may arise from incidental text rewriting by GPT-4o, not from emoji substitution.","rationale":"The reader's weakest assumption (unvalidated GPT-4 role labels) is real but secondary: even if role labels are noisy, the pipeline could still lower offensiveness by replacing emojis in a generally safer direction. The more fundamental threat is that the pipeline may not be emoji-only at all. The Step 3 prompt (Appendix B.2) requests a 'full tweet after all necessary replacements' and never prohibits editing non-emoji tokens; the human evaluation (Section 4.5) reports only aggregate scores, and the only manipulation check (Section 4.4) is emoji-frequency based. Without a token-level diff, the measured offensiveness drop and the semantic-preservation numbers could reflect text rewriting, which would invalidate the paper's core contribution: emoji-specific moderation. This is a concrete, fixable omission: an edit-distance analysis or a surgical-replacement experiment would settle it. I therefore agree with the CONDITIONAL verdict, but for a reason partly different from the reader's.","tokens_in":16641,"tokens_out":8550,"duration_ms":80059,"concrete_test":"Extract the 200 original and 200 multi-step revised tweets from Section 4.5. Remove emoji characters from both versions (using a Unicode emoji segmentation library), then compute word-level and character-level edit distance. If any non-emoji token differs in more than 1% of pairs, the manipulation is not emoji-only. To test causality, construct a 'surgical' condition by applying only the emoji substitutions chosen by Step 3 to the original tweet, leaving every other byte identical, and collect offensiveness ratings for these surgical versions. If the surgical versions do not reproduce the full pipeline's offensiveness reduction (e.g., the mean drop is significantly smaller than 0.42), the headline effect is confounded by text rewriting.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—that selectively replacing harmful emojis reduces offensiveness—requires that the pipeline changes only emojis. The paper asserts it 'target[s] only the emojis likely to cause harm' (Section 4), but the Step 3 prompt in Appendix B.2 asks the LLM to provide 'the full tweet after all necessary replacements' without any instruction to preserve non-emoji tokens verbatim. The human evaluation (Section 4.5) compares original and processed tweets, but the only manipulation check reported (Section 4.4, Table 9) concerns emoji frequency distributions; no token-level diff or edit-distance analysis is provided. It is therefore possible that GPT-4o also softens profanity, removes insults, or rephrases clauses, in which case the observed offensiveness reduction (3.00 → 2.58, p<0.05) would be at least partially attributable to text rewriting rather than to the proposed emoji-replacement mechanism. The direct-prompting baseline has the same ambiguity, so the reported superiority of the multi-step pipeline could also reflect different amounts of textual revision. Because the paper's contribution is specifically emoji moderation, this unverified assumption is load-bearing: without a demonstration that only emojis changed, the human evaluation cannot support the claimed mechanism.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the role of emojis in offensive tweets and proposes a multi-step LLM pipeline that selectively replaces emojis judged to be offensive in themselves or to intensify offense, while preserving the tweet's semantics. Using a collected dataset of 9,285 offensive tweets, the authors analyze emoji roles and distributions across offense types, then run a human evaluation (20 annotators, 200 tweets in three conditions) comparing original tweets, pipeline-processed tweets, and a direct-prompting baseline. The headline result is that the pipeline reduces perceived offensiveness from 3.00 to 2.58 on a 1–5 scale with p<0.05, with no notable changes in semantic measures. The paper also reports heterogeneous effects across offense types and provides case studies and prompts in the appendix.","tokens_in":16858,"tokens_out":5541,"duration_ms":48898,"significance":"If the main result holds, the paper is a useful contribution to an under-studied area: it provides descriptive evidence that emojis have distinct functional roles in offensive content and proposes a more conservative moderation alternative than full-tweet rewriting. The descriptive findings on emoji distributions by offense type and role (Figures 1 and Tables 1–3) are valuable for future research on emoji semantics and content moderation. However, the central causal claim—that the offensiveness reduction is specifically due to emoji substitution—is not yet verified, because the pipeline does not demonstrably leave non-emoji text untouched. In addition, the statistical reporting is too thin to support the strength of the conclusions, and the paper does not release data or code, which limits reproducibility. These issues are fixable, but they are load-bearing for the paper's main contribution.","major_comments":[{"comment":"The claim that the pipeline reduces offensiveness through emoji substitution alone is not supported, because the Step 3 prompt instructs the LLM to output 'the full tweet after all necessary replacements' without requiring that non-emoji tokens be preserved verbatim, and the direct-prompting baseline has the same ambiguity. The only manipulation check in Section 4.4 (Table 9) compares emoji frequency distributions; no token-level diff or edit-distance analysis is reported. Consequently, the offensiveness drop from 3.00 to 2.58 in Table 5 could be partially or wholly caused by incidental text rewriting (e.g., softening profanity or rephrasing), not by emoji replacement. Because the paper's stated contribution is targeted emoji moderation, this missing control is load-bearing; the authors should either constrain the prompt to forbid non-emoji edits and verify with a diff, or add a text-only rewriting condition to isolate the emoji effect.","section":"Section 4.2, Step 3; Appendix B.2"},{"comment":"The statistical evidence for the headline result is underreported. The paper reports only mean scores and a 'p<0.05 (paired t-test)' asterisk, with no standard deviations, confidence intervals, effect sizes, or inter-annotator agreement (e.g., Krippendorff's alpha or Cohen's kappa) despite each tweet being rated by two annotators. The analysis also lacks any correction for multiple comparisons: Table 6 tests five offense types across roughly ten variables (about 50 pairwise tests), so the significant offensiveness deltas (racial, sexual, violence) and the body-symbol/dehumanization/extra-meaning deltas are expected by chance at the 0.05 level. Without these details, the claim that the pipeline 'significantly reduces offensiveness' is not adequately established.","section":"Section 4.5, Tables 5 and 6"},{"comment":"The GPT-4-based emoji role labels are used both as the descriptive analysis (Figure 1, Table 1) and as prior knowledge injected into the pipeline prompt (the lists of 'Emojis Often Associated with Specific Offense Types' and 'Popular Emojis by Typical Functionality'). The paper never validates these labels against human judgments, so the pipeline may be systematically replacing or retaining emojis based on noisy or biased role assignments. This matters because Step 2 determines which emojis are candidates for replacement; if the role labels are unreliable, the observed offensiveness reduction may not generalize to other datasets or models. A small human-annotation reliability study for the four-way role taxonomy would address this.","section":"Section 3.1 and Section 4.2, Step 2"},{"comment":"The within-subject design presents the same tweet in original, pipeline-processed, and baseline-processed forms to the same annotators. Although the paper randomizes presentation order, it does not describe counterbalancing or whether annotators could recognize repeated content; repeated exposure may induce demand characteristics and inflate the apparent effect. At minimum, the authors should report whether they used a Latin-square design or between-subject split, and analyze order effects.","section":"Section 4.5, Evaluation Design"}],"minor_comments":[{"comment":"These tables are empty in the manuscript text as provided because the emoji glyphs are missing, so the reader cannot verify the actual emoji lists; please include Unicode names or a rendered figure.","section":"Tables 1, 2, 3, and 9"},{"comment":"The thresholds for 'High/Moderate/Low frequency' (30% and 20%) are arbitrary and not justified; since these categories are later fed into the Step 2 prompt, the threshold choice is a free parameter that should be justified or examined in a sensitivity analysis.","section":"Section 3.3, Table 3"},{"comment":"The description of the annotation procedure is ambiguous: 'Each annotator assesses 60 tweets ... consisting of 20 original tweets, 20 versions ...' could mean each annotator sees triples of the same 20 tweets, or 60 distinct items; please clarify the assignment and the unit of the paired t-test (tweet-level or annotator-level).","section":"Section 4.5, Experimental Setup"},{"comment":"The question 'Do you agree ... meanings of all emojis ... disclosed by the text' uses a two-point Disagree/Agree scale; the 'Extra Meaning (% Yes)' in Table 5 presumably reflects agreement, but the mapping is not stated.","section":"Appendix A.2, Extra Meaning question"},{"comment":"The Limitations section acknowledges cultural and LLM-bias concerns but does not mention the possibility that the LLM rewrites non-emoji text; adding this limitation and a corresponding analysis would strengthen the paper.","section":"Section 7, Limitations"},{"comment":"Several references lack arXiv IDs or DOIs (e.g., Davidson et al. 2019, Barbieri et al. 2020); please standardize the bibliography.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about text rewriting is the key issue and, on reading the paper, it does land: the Step 3 prompt and the direct baseline do not constrain non-emoji tokens, and the manipulation check is only emoji-frequency based. Please ask the authors to add a token-level diff analysis or constrain the prompt and verify that only emoji tokens change. Also flag the absence of inter-annotator agreement and multiple-comparison corrections as essential statistical fixes."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the descriptive half of this paper is a solid, useful contribution: a taxonomy of emoji roles (offense itself, intensifier, mitigator, unrelated) applied to 9,285 offensive tweets, with distributions across offense types and frequency analysis. That part is worth reading and citing. Second, the moderation pipeline's central claim—that selectively replacing emojis lowers offensiveness—does not yet have evidence that only emojis changed. The Step 3 prompt asks the LLM for \"the full tweet after all necessary replacements\" and never instructs it to preserve non-emoji tokens. The paper reports emoji-frequency checks but no token-level diff. So the 3.00→2.58 offensiveness drop could be partly from incidental text rewriting. The stress-test note lands.\n\nNow the detail. The paper's novelty is real: prior work separates emoji sentiment from offensive-text detection, and this is the first systematic look at emoji function in offensive content. The GPT-4-based role labeling is plausible but unvalidated against human judgments; that matters for the analysis claims, though the main human evaluation of the pipeline is independent of the LLM labels. The human evaluation itself is small: 20 annotators, 200 tweets per condition, no inter-annotator agreement, no confidence intervals, no multiple-comparison correction. The significant overall effect is modest and driven by racial, sexual, and violence categories; personal attacks and political tweets show no significant reduction, which the authors honestly report. The semantic preservation claim rests on null results from the same small sample, so it is weak.\n\nThe manipulation-check gap is the biggest problem. If the pipeline also rewrites text, the paper's framing as 'emoji replacement' is misleading. The fix is straightforward: constrain the prompt to change only emojis, and report edit distance or token-level diffs. This is a revision, not a rejection.\n\nWho should read it: anyone working on emoji semantics, content moderation, or LLM-based rewriting. It deserves a serious referee because the descriptive analysis is a genuine addition and the pipeline question is worth answering properly. I'd bring it to a reading group, and I'd cite the taxonomy. My recommendation: engage, but push for the manipulation check and reporting fixes.","headline":"Useful descriptive analysis of emoji roles in offensive tweets, but the pipeline's claimed emoji-only mechanism is unverified because the prompt allows text rewriting.","tokens_in":17413,"tokens_out":2923,"would_cite":true,"duration_ms":27095,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Emoji swaps lower tweet offensiveness without changing meaning, human raters say","keywords":["emoji semantics","offensive content detection","content moderation","large language models","GPT-4","Twitter","sarcasm","offense taxonomy"],"falsifier":"Take a random sample of the 9,285 offensive tweets, ask a group of human annotators to label every emoji with the paper's four-role taxonomy, and measure agreement with GPT-4's Step-2 labels. If agreement is near chance, the pipeline is targeting the wrong emojis and its observed offensiveness drop would not generalize. A second falsifier: run the pipeline only on tweets human-raters scored 5 (extremely offensive); the paper itself reports no offensiveness reduction there, so the central claim should be scoped to mild and moderate offenses, and a replication restricted to score-5 tweets would be expected to show zero effect.","tokens_in":16400,"feed_emoji":"🤬","tokens_out":3605,"duration_ms":33137,"temperature":0.7,"pith_summary":"This paper claims that emojis are a systematic, understudied carrier of offense in tweets: they can be offensive in themselves, intensify an attack, soften it, or be unrelated, and their role shifts with context and offense type. It then builds a four-step, GPT-4-powered pipeline that classifies each emoji's role and replaces only those that are offensive or intensifying, rather than rewriting the entire tweet. Human evaluation of 200 tweets shows the approach cuts perceived offensiveness from 3.00 to 2.58 on a 1-5 scale with p<0.05, while semantic measures (sentiment, arousal, clarity, fluency, extra meaning) stay essentially unchanged. The paper argues that content moderation should therefore treat emojis as meaningful contributors to harm, not decoration. The effect is heterogeneous: it is strongest for mildly offensive tweets and weak or absent when strong verbal abuse already dominates.","feed_headline":"Emoji swaps cut tweet offensiveness without changing meaning","feed_subtitle":"Replacing only harmful emojis, not the whole text, lowers offense scores from 3.00 to 2.58 in human ratings.","key_machinery":"The load-bearing machinery is a four-role taxonomy of emoji function in offensive tweets—offensive in itself, intensifies offense, mitigates offense, unrelated to offense—paired with a four-step LLM pipeline. The pipeline first classifies whether a tweet is offensive and which of five offense types it belongs to, then assigns every emoji a role using the taxonomy and prior distributional findings, then recommends surrogate emojis only for emojis labeled offensive-in-themselves or intensifying, and finally condenses the justifications into a user-facing summary. The role taxonomy is what makes replacement selective rather than destructive: it directs the LLM to leave mitigating and unrelated emojis untouched, which is why semantic measures remain stable while offensiveness falls.","core_discovery":"Emojis carry offense through at least four distinct roles—being the offense itself, intensifying the offense, mitigating it, or being unrelated to it—and these roles are predictable enough from context that a prompted LLM can use them to selectively replace harmful emojis. In a within-subject human experiment, the pipeline's targeted replacements reduced perceived offensiveness from 3.00 to 2.58 (p<0.05) with no statistically significant change in sentiment, arousal, clarity, or fluency. A direct-prompting baseline that simply asks the LLM to replace offensive emojis without the role taxonomy produced only a small, insignificant drop (3.00 to 2.94), indicating that the taxonomy-driven stepwise procedure is what carries the effect. The paper also documents that emoji-offense associations are type-specific: body-part emojis cluster in sexual content, dehumanizing emojis in personal and political attacks, and the pipeline's substitutions successfully reduce those specific markers even when the overall offensiveness score does not move. Removing emojis only changes perceived offensiveness for mildly offensive tweets; when verbal abuse alone earns a score of 4 or 5, emoji replacement has little effect.","pith_inferences":["The entire pipeline hinges on GPT-4's Step-2 role labels, and since the paper never validates those labels against human judgments, the 3.00-to-2.58 effect could be an artifact of the model's own consistent but incorrect role assignments; a human-annotation check of emoji roles on a random sample would settle this.","Because the paper's own limitation section notes cultural and individual variation in emoji interpretation, the pipeline is likely to be far less effective outside U.S. English contexts; a natural extension is to build role taxonomies per language or culture rather than one global taxonomy.","The finding that emoji replacement works only for mild offense suggests a deployment strategy: platforms could use this pipeline as a low-stakes nudge for borderline tweets while leaving high-severity content to the ordinary removal or flagging process.","The paper's taxonomy could be imported into hate-speech datasets to create emoji-role annotations at scale, which would let future work train smaller, faster classifiers to do the replacement without calling an LLM for every tweet."],"forward_implications":["Content moderation systems should treat emojis as first-class carriers of offensive meaning rather than as noise, and can leverage role taxonomies to target them selectively.","An LLM given only a generic instruction to reduce offensiveness through emoji replacement fails to lower offense scores, whereas the same model guided by an explicit role taxonomy succeeds, showing that prompt structure, not model capability alone, drives the outcome.","Emoji moderation is most effective for mildly offensive content; for tweets already rated 4-5 on offensiveness, replacing emojis does not move the needle, so deployment should prioritize borderline cases.","The effect is offense-type specific: the pipeline reduces dehumanizing emojis in personal and political attacks (by 12.8% and 22.5%) and body-symbol emojis in sexual content (by 15.0%), without inflating sarcasm.","Replacing emojis that carry implicit offense (body symbols, dehumanizing objects, sarcastic faces) can preserve the semantic intent of a tweet while lowering its perceived harm, offering a less invasive alternative to full-text rewriting."],"supporting_citations":[{"why":"Provides the definition of offensive content (unacceptable language or targeted offense, direct or veiled) that the paper adopts for labeling and evaluation.","marker":"(Poletto et al., 2021)"},{"why":"Supplies a standard offensive-language taxonomy and classification setup that the paper builds on for offense detection and type taxonomy.","marker":"(Zampieri et al., 2019)"},{"why":"The RoBERTa model (twitter-roberta-base-offensive) that filters candidate offensive tweets from the collected corpus before GPT-4 labeling.","marker":"(Liu et al., 2019; Barbieri et al., 2020)"},{"why":"The GPT-4 / GPT-4o model used for offensive-content classification, emoji role labeling, surrogate recommendation, and justification generation.","marker":"(OpenAI, 2023)"},{"why":"BERTopic, the unsupervised topic-modeling method used to cluster tweets into the 14 offensive topics later grouped into five offense types.","marker":"(Grootendorst, 2022)"},{"why":"Empirical evidence that emojis can carry irony and sarcasm, grounding the taxonomy's 'intensifies offense' role.","marker":"(Weissman and Tanner, 2018)"},{"why":"Documents sender-intended functions of emojis, supporting the 'mitigates offense' role in the taxonomy.","marker":"(Cramer et al., 2016)"},{"why":"Shows how emoji sequences are used in enacting identity and adjusting tone, cited for the mitigate-function of emojis.","marker":"(Ge, 2019)"},{"why":"Provides background on hate speech and offensive-language datasets and on racial bias in detection, framing the paper's motivation to move beyond text-only detection.","marker":"(Davidson et al., 2019)"}],"fun_headline_variants":["Emoji role taxonomy drives 14% offense cut","Targeted emoji replacement reduces tweet offensiveness","LLM swaps harmful emojis, preserves meaning","Emoji offense roles exposed, pipeline defuses tweets","Role-aware emoji moderation beats blunt prompting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pipeline's second step assumes that GPT-4 can reliably tell, from the tweet text alone, which emoji is offensive by itself, which intensifies the attack, which softens it, and which is irrelevant, but the paper never checks these role labels against human judgment.","fun_headline_variants_meta":{"raw":{"variants":["Emoji role taxonomy drives 14% offense cut","Targeted emoji replacement reduces tweet offensiveness","LLM swaps harmful emojis, preserves meaning","Emoji offense roles exposed, pipeline defuses tweets","Role-aware emoji moderation beats blunt prompting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000684,"raw_usage":{"total_tokens":3110,"prompt_tokens":956,"completion_tokens":2154,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":572,"completion_tokens_details":{"reasoning_tokens":2080}},"tokens_in":572,"tokens_out":2154,"duration_ms":15329,"temperature":1.0,"reasoning_tokens":2080,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:02:14.672546+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a random sample of the 9,285 offensive tweets, ask a group of human annotators to label every emoji with the paper's four-role taxonomy, and measure agreement with GPT-4's Step-2 labels. If agreement is near chance, the pipeline is targeting the wrong emojis and its observed offensiveness drop would not generalize. A second falsifier: run the pipeline only on tweets human-raters scored 5 (extremely offensive); the paper itself reports no offensiveness reduction there, so the central claim should be scoped to mild and moderate offenses, and a replication restricted to score-5 tweets would be expected to show zero effect.","supporting_citations":[],"review_version":1}