Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Moderating Harm: Benchmarking Large Language Models for Cyberbullying Detection in YouTube Comments

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Across 5,080 real YouTube comments, GPT-4.1 gets the best zero-shot cyberbullying detection balance, with F1 0.863.

desk verdict Useful new dataset and honest error analysis; the model ranking is prompt-dependent and stated too strongly. read the letter →

arxiv 2505.18927 v3 pith:UAXPJ7Y7 submitted 2025-05-25 cs.CL cs.AI

Amel Muminovic This is my paper
classification cs.CLcs.AI
keywords cyberbullyingdetectioncontentmoderationlargelanguagemodelszero-shotclassificationYouTubecommentshatespeechmultilingualabuseGPT-4.1
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 benchmarks three large language models—OpenAI's GPT-4.1, Google's Gemini 1.5 Pro, and Anthropic's Claude 3 Opus—as zero-shot detectors of cyberbullying in a corpus of 5,080 YouTube comments from high-abuse threads in gaming, lifestyle, food-vlog, and music videos. Two annotators independently labeled each comment (Cohen's $\kappa = 0.83$), and all three models saw the same minimal prompt under deterministic decoding. GPT-4.1 achieves the best overall balance, with F1 0.863, precision 0.887, and recall 0.841. Gemini catches the most harmful comments (recall 0.875) but over-flags benign ones (precision 0.767), while Claude is the opposite: the most precise (0.920) with the fewest false positives (0.022) but the lowest recall (0.720). The paper's central message is that no single model fits every moderation policy, and its qualitative review shows all three miss sarcasm, coded insults, and mixed-language slang; the de-identified data, prompts, and outputs are released so the comparison can be reproduced and extended.

What carries the argument

The controlled comparison mechanism is a fixed zero-shot protocol: every comment is classified with the same minimal instruction ('Analyze the following comment and determine if it contains cyberbullying, harassment, or harmful language... Respond with 0 if it is not harmful and 1 if it is harmful'), with temperature set to 0.0 and max_tokens capped at 10. That single prompt turns the three models into the only variable and is what allows precision, recall, and F1 differences to be attributed to the models' own zero-shot moderation policies. The dataset construction carries part of the argument too: pilot crawls kept only videos whose comment sections had more than 20% abuse, the final 5,080 comments were labeled independently by two annotators with Cohen's $\kappa = 0.83$ and adjudicated to consensus, so the ground truth is designed to contain the subtle, context-dependent harm that existing isolated-snippet benchmarks omit.

What would settle it

Run the same 5,080 comments through several differently worded versions of the prompt—defining cyberbullying, asking for a rationale before the label, or adding few-shot examples—under the same deterministic settings; if GPT-4.1 does not keep the highest F1, the reported ranking is an artifact of the single prompt rather than a stable property of the models.

Watch

Extended reading notes

Core claim

On 5,080 authentic YouTube comments drawn from four high-abuse threads, GPT-4.1, Gemini 1.5 Pro, and Claude 3 Opus, queried with an identical minimal prompt and temperature 0.0, produce three clearly different moderation profiles. GPT-4.1 delivers the best balance, with F1 0.863, precision 0.887, and recall 0.841. Gemini 1.5 Pro finds the most harmful content (recall 0.875) but at the cost of the most false positives (precision 0.767, false-positive rate 0.095). Claude 3 Opus is the most cautious: precision 0.920, false-positive rate 0.022, recall 0.720. The qualitative review of 200 disagreements describes the blind spots behind these numbers—sarcasm, mock praise, faux concern, emoji-driven ridicule, and collective piling-on—and the paper concludes that moderation pipelines should pair complementary models, add thread-level context, and tune for under-represented languages and implicit abuse rather than rely on any one zero-shot system.

Load-bearing premise

The headline ordering rests on the untested assumption that one fixed prompt, used verbatim for all three models, is a fair and representative test of each model's moderation ability; no alternative prompt is tried.

Editorial extensions

If this is right

  • A platform that wants one out-of-the-box zero-shot filter can expect GPT-4.1 to catch roughly 84% of harmful comments in this kind of YouTube thread while keeping precision near 0.89.
  • A safety-first platform can deploy Gemini as a high-recall first pass and route its flags to a high-precision check, since Gemini surfaces the most true positives and Claude removes most of Gemini's false positives at the second stage.
  • A platform that treats false accusations as costly would prefer Claude, whose false-positive rate of 0.022 means almost every flag is defensible, but it must accept that only 72% of harmful comments will be caught.
  • Comment-by-comment moderation, regardless of model, will miss sarcasm, mock praise, faux concern, and coordinated piling-on; the paper's qualitative examples show these errors persist across all three systems.
  • The released de-identified corpus and full prompts give other researchers a fixed YouTube ground-truth set on which to compare future models and moderation pipelines against the published scores.

Reading between the lines

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

  • Inference: Because only one prompt is tested per model, the headline ranking is a property of those exact instructions, not of the models' ceilings; a small prompt-sweep could reorder GPT-4.1, Gemini, and Claude.
  • Inference: The released per-comment outputs make a direct test of the proposed two-stage pipeline possible—run Gemini as the recall filter, ask Claude to re-score only Gemini's flags, and compare the ensemble F1 against GPT-4.1's single-model score.
  • Inference: Because all 5,080 comments come from only four deliberately abusive threads, the 26% harm rate is far higher than a typical YouTube comment section; under more realistic, lower-prevalence conditions the false-positive costs grow, which would likely favor Claude's conservative profile.
  • Inference: The qualitative error themes suggest a targeted extension: a sub-benchmark isolating sarcasm, coded insults, and code-switched slang could quantify how much of each model's remaining error budget is implicit rather than explicit harm, and would be a sharper test than overall F1.
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 benchmarks three LLMs — OpenAI GPT-4.1, Google Gemini 1.5 Pro, and Anthropic Claude 3 Opus — for zero-shot cyberbullying detection on 5,080 YouTube comments drawn from four videos in gaming, lifestyle, food-vlog, and music channels. The corpus contains 1,334 harmful and 3,746 non-harmful messages in English, Arabic, and Indonesian, labeled independently by two annotators with reported Cohen's kappa of 0.83. All models are evaluated with one identical minimal prompt under temperature 0.0. GPT-4.1 achieves the best reported balance (F1 = 0.863, precision = 0.887, recall = 0.841), Gemini the highest recall (0.875) but lower precision (0.767), and Claude the highest precision (0.920) and lowest false-positive rate (0.022) but reduced recall (0.720). The authors also present qualitative error analyses, a comparison with a published BERT baseline, and release the de-identified corpus, prompts, and model outputs.

Significance. If the result holds in the qualified form the data actually support, the paper is a useful, reproducible snapshot of current zero-shot LLM moderation behavior on authentic YouTube content. Strengths include deterministic decoding, a fixed prompt, an openly released corpus and outputs, and internally consistent count-to-metric calculations (Tables II and III agree with the stated formulas). The absence of fitted parameters and the use of independent human labels make the benchmark itself non-circular. The main significance risk is that the headline ranking is presented as a model-level property despite resting on one arbitrary prompt and only four source videos; the paper's own related-work section and limitations acknowledge prompt sensitivity, so the claim needs substantial qualification or additional robustness evidence.

major comments (3)
  1. [Section III-E, Tables II-III, Abstract] The central claim that GPT-4.1 is the 'best overall' model is established under exactly one hand-written zero-shot prompt. The paper itself cites evidence in Section II-D that LLM responses are highly sensitive to prompt phrasing, and the Limitations section concedes that the zero-shot setup 'may understate each model's peak performance.' With no prompt-variant sensitivity tests, the observed F1 gaps (0.863 vs. 0.818 vs. 0.808) could reflect which model happens to match this particular wording rather than a stable model-level ordering. Because the prompts and outputs are released, this is directly testable. I recommend either adding prompt-variant experiments (e.g., several paraphrases with and without label definitions) and showing that the ordering is stable, or consistently qualifying all 'best' claims as 'best under the single prompt used here.' As written, the Abstract and Conclusion overstate the generality of the ranking.
  2. [Section III-A, Table III] The dataset is drawn from only four videos, selected by a pilot abuse rate above 20%, and no per-video or per-domain statistics are reported. Selection on the outcome variable, combined with the very small number of source videos, means that the pooled F1 differences and the qualitative error themes may be specific to these particular threads rather than representative of the four content domains named in the title and introduction. At minimum, the authors should report per-video and per-domain sample sizes and metric breakdowns, and add paired significance tests (e.g., McNemar) for the GPT-vs-Gemini and GPT-vs-Claude differences in Table III. Without uncertainty estimates, the reader cannot tell whether the reported ordering is statistically meaningful.
  3. [Section III-C, Limitations #3] The ground-truth labels are load-bearing for every result in the paper, yet they come from two annotators with 'similar linguistic and cultural backgrounds,' one of whom is the first author. The paper acknowledges this risk in the Limitations but does not quantify it. I recommend reporting inter-annotator agreement separately for English, Arabic, and Indonesian and for each of the four video domains, and describing the adjudication process in more detail (e.g., how often discussion changed an initial label, and whether disagreements clustered in the sarcasm/coded-language categories that the paper identifies as the models' main failure modes). This would materially strengthen the dataset contribution.
minor comments (6)
  1. [Abstract vs. Section III-C] The front-matter abstract describes kappa = 0.83 as 'substantial agreement,' while the full-text abstract and Section III-C call it 'almost perfect.' Under Landis and Koch's categories, 0.83 is 'almost perfect'; the wording should be made consistent.
  2. [Section IV-C, Limitations #5] Table IV reports a fine-tuned BERT baseline, but the Limitations section states 'No BERT-based baseline.' This is internally contradictory. Either remove the table or revise the Limitations item; the table should also be labeled explicitly as an illustrative, cross-corpus reference point rather than a 'baseline comparison,' since the BERT score is on HateXplain while the LLM scores are on the new YouTube corpus.
  3. [Table IV] The Macro-F1 values in Table IV for the three LLMs are identical to the harmful-class F1 values in Table III. With a 26%/74% class split, macro-averaged F1 cannot equal the harmful-class F1 unless the non-harmful class has the same F1, which the confusion matrices in Table II show is not the case. Please report the actual macro-F1 calculations or relabel the column.
  4. [Abstract and Section III-A] The paper calls the evaluation a 'multi-language comparison,' but Limitations #4 states the dataset is 'overwhelmingly English.' Recommend changing the abstract to 'English-dominant corpus with Arabic and Indonesian samples' to avoid overstating language coverage.
  5. [Section I-B] The phrase 'three most advanced LLMs' is subjective and time-sensitive; 'three widely used LLMs' would be more precise and less likely to date the paper. Similar wording appears in the Conclusion.
  6. [Data Availability] The repository URL contains a literal space encoded as '%20'; ensure the published URL is clean and stable, and consider adding a DOI or versioned release to support the reproducibility claim.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: model rankings are measured against independent human labels, not constructed from the inputs.

full rationale

The paper's central claim—that GPT-4.1 achieves the best F1 balance (0.863) among three zero-shot LLMs—is an empirical measurement, not a derivation. Each model is queried with an identical fixed prompt at temperature 0.0, and the outputs are compared directly to 5,080 independently annotated human labels (κ = 0.83). No parameter is fitted to any subset of the target data, no prompt is optimized on the test comments, and no equation defines the reported F1, precision, or recall in terms of the model identity or the rubric. The claimed ranking is therefore statistically forced only by the actual model responses, not by construction. The single-prompt design is a genuine validity limitation, honestly stated in Limitations item 1 ('may understate each model's peak performance'), but that is a robustness concern about prompt sensitivity, not circularity, since the numbers remain measured outcomes rather than entailed consequences of the design. The pilot-crawl selection rule (retaining videos with pilot abuse rate above 20%) influences generalizability but does not encode the target results. The BERT baseline comparison is explicitly flagged as not a head-to-head benchmark because the corpora differ. There are no load-bearing self-citations, no uniqueness theorems, and no renamed known results. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the quality and representativeness of the human labels, the selected videos, and the single prompt. No new entities, particles, or forces are introduced. The pilot threshold is a design choice rather than a fitted model parameter.

free parameters (1)
  • Video inclusion threshold (pilot abuse rate) = >20%
    Chosen by hand to retain only high-abuse videos for the corpus (Section III-A). It shapes dataset prevalence and generalizability but is not fitted to model outputs.
assumptions (4)
  • domain assumption Ground-truth labels from two reviewers accurately capture harmful intent.
    Section III-C reports high agreement (kappa 0.83) but both annotators share similar cultural backgrounds, acknowledged in Limitations 3, so label bias is possible.
  • domain assumption Four selected videos are representative of cyberbullying on YouTube.
    Section III-A uses only four videos retained after a pilot abuse-rate filter, and video IDs are withheld, so external validity is limited.
  • domain assumption A single fixed zero-shot prompt provides a fair cross-model comparison.
    Section III-E describes one prompt with no variation or robustness test, although LLM rankings are known to be prompt-sensitive.
  • domain assumption Binary harmful/not-harmful labeling adequately represents cyberbullying for this task.
    Section III-C collapses degrees and types of harm into one class, which may affect how precision and recall should be interpreted.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Moderating Harm: Benchmarking Large Language Models for Cyberbullying Detection in YouTube Comments." pith.science (2026). https://pith.science/paper/UAXPJ7Y7

@misc{pith2026250518927,
  author       = {Pith},
  title        = {Pith review of: Moderating Harm: Benchmarking Large Language Models for Cyberbullying Detection in YouTube Comments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UAXPJ7Y7}},
  note         = {Machine review of arXiv:2505.18927}
}
read the original abstract

As online platforms grow, comment sections increasingly host harassment that undermines user experience and well-being. This study benchmarks three leading large language models, OpenAI GPT-4.1, Google Gemini 1.5 Pro, and Anthropic Claude 3 Opus, on a corpus of 5,080 YouTube comments sampled from high-abuse threads in gaming, lifestyle, food vlog, and music channels. The dataset comprises 1,334 harmful and 3,746 non-harmful messages in English, Arabic, and Indonesian, annotated independently by two reviewers with substantial agreement (Cohen's kappa = 0.83). Using a unified prompt and deterministic settings, GPT-4.1 achieved the best overall balance with an F1 score of 0.863, precision of 0.887, and recall of 0.841. Gemini flagged the highest share of harmful posts (recall = 0.875) but its precision fell to 0.767 due to frequent false positives. Claude delivered the highest precision at 0.920 and the lowest false-positive rate of 0.022, yet its recall dropped to 0.720. Qualitative analysis showed that all three models struggle with sarcasm, coded insults, and mixed-language slang. These results underscore the need for moderation pipelines that combine complementary models, incorporate conversational context, and fine-tune for under-represented languages and implicit abuse. A de-identified version of the dataset and full prompts is publicly released to promote reproducibility and further progress in automated content moderation.

Figures

Figures reproduced from arXiv: 2505.18927 by the authors.

Figure 1
Figure 1. Comparison of precision, recall, and F1 score across evaluated models. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Large Language Models for Toxic Language Detection in Low-Resource Balkan Languages

    cs.CL 2025-06 conditional novelty 5.0 of 10

    Adding a brief, hand-written video context improved LLM toxicity detection in Serbian, Croatian, and Bosnian, raising F1 scores by up to 0.10 across four models.

Reference graph

Works this paper leans on

44 extracted references · 17 canonical work pages · cited by 1 Pith paper

  1. [1]

    Social media usage & growth statistics,

    B. Dean, “Social media usage & growth statistics,” Backlinko, Feb. 21,

    work page

  2. [2]

    Study of cyberbullying among adoles- cents in recent years: A bibliometric analysis,

    A. B. Barrag ´an Mart ´ın et al. , “Study of cyberbullying among adoles- cents in recent years: A bibliometric analysis,” Int. J. Environ. Res. Public Health, vol. 18, no. 6, p. 3016, Mar. 2021. [Online]. Available: https://doi.org/10.3390/ijerph18063016

    work page doi:10.3390/ijerph18063016 2021

  3. [3]

    Bullying, cyberbullying, and sui- cide,

    S. Hinduja and J. W. Patchin, “Bullying, cyberbullying, and sui- cide,” Arch. Suicide Res. , vol. 14, no. 3, pp. 206–221, 2010. doi:10.1080/13811118.2010.494133

    arXiv 2010

  4. [4]

    Anonymously hurting others online: The effect of anonymity on cyberbullying frequency,

    C. P. Barlett, “Anonymously hurting others online: The effect of anonymity on cyberbullying frequency,” Psychol. Pop. Media Cult., vol. 4, no. 2, pp. 70–79, 2015. doi:10.1037/a0034335

    work page doi:10.1037/a0034335 2015

  5. [5]

    The severity of cyberbullying affects bystander inter- vention among college students: The roles of feelings of responsibility and empathy,

    L. Huang et al., “The severity of cyberbullying affects bystander inter- vention among college students: The roles of feelings of responsibility and empathy,” Psychol. Res. Behav. Manag., vol. 16, pp. 893–903, Mar

    work page

  6. [6]

    Cyberbullying: Twenty crucial statistics for 2024,

    A. Vigderman, “Cyberbullying: Twenty crucial statistics for 2024,” Security.org, Oct. 9, 2024. [Online]. Available: https://www.security.org/resources/cyberbullying-facts-statistics/

    work page 2024

  7. [7]

    Social media use and cyber-bullying: A cross-national analysis of young people in 42 countries,

    W. Craig et al., “Social media use and cyber-bullying: A cross-national analysis of young people in 42 countries,” J. Adolesc. Health , vol. 66, no. 6, pp. S100–S108, Jun. 2020. doi:10.1016/j.jadohealth.2020.03.006

    work page doi:10.1016/j.jadohealth.2020.03.006 2020

  8. [8]

    Automated content moderation increases adherence to community guidelines,

    M. H. Ribeiro, J. Cheng, and R. West, “Automated content moderation increases adherence to community guidelines,” in Proc. ACM Web Conf. (WWW), 2023, pp. 2666–2676. doi:10.1145/3543507.3583275

    arXiv 2023

Show all 44 references
  1. [9]

    Content moderation on social media: Does it matter who and why moderates hate speech?

    S. Wang and K. J. Kim, “Content moderation on social media: Does it matter who and why moderates hate speech?” Cyberpsy- chol. Behav. Soc. Netw. , vol. 26, no. 7, pp. 527–534, Jul. 2023. doi:10.1089/cyber.2022.0158

    2023

  2. [10]

    Content moderation, AI, and the question of scale,

    T. Gillespie, “Content moderation, AI, and the question of scale,” Big Data Soc. , vol. 7, no. 2, pp. 1–5, Jul. 2020. doi:10.1177/2053951720943234

    2020 doi

  3. [11]

    Context in abusive language detection: On the interdependence of context and annotation of user comments,

    H. Lopez and S. K ¨ubler, “Context in abusive language detection: On the interdependence of context and annotation of user comments,” Discourse, Context Media , vol. 63, Art. no. 100848, Feb. 2025. doi:10.1016/j.dcm.2024.100848

    2025

  4. [12]

    Relationship between peer victimization, cyberbullying, and suicide in children and adolescents: A meta-analysis,

    M. van Geel, P. Vedder, and J. Tanilon, “Relationship between peer victimization, cyberbullying, and suicide in children and adolescents: A meta-analysis,” JAMA Pediatr., vol. 168, no. 5, pp. 435–442, May 2014. doi:10.1001/jamapediatrics.2013.4143

    2014

  5. [13]

    Hateful symbols or hateful people? Predic- tive features for hate speech detection on Twitter,

    Z. Waseem and D. Hovy, “Hateful symbols or hateful people? Predic- tive features for hate speech detection on Twitter,” in Proc. NAACL Student Res. Workshop , San Diego, CA, USA, Jun. 2016, pp. 88–93. doi:10.18653/v1/N16-2013

    2016 doi

  6. [14]

    Large scale crowdsourcing and characterization of Twitter abusive behavior,

    A.-M. Founta et al., “Large scale crowdsourcing and characterization of Twitter abusive behavior,” inProc. Int. Conf. Web Social Media, Atlanta, GA, USA, Mar. 2018, pp. 491–500. doi:10.1609/icwsm.v12i1.14991

    2018 doi

  7. [15]

    Predicting the type and target of offensive posts in social media,

    M. Zampieri et al., “Predicting the type and target of offensive posts in social media,” in Proc. NAACL, Minneapolis, MN, USA, Jun. 2019, pp. 1415–1420. doi:10.18653/v1/N19-1144

    2019 doi

  8. [16]

    HateCheck: Functional tests for hate speech detection models,

    P. R ¨ottger, B. Vidgen, D. Nguyen, Z. Waseem, H. Margetts, and J. Pierrehumbert, “HateCheck: Functional tests for hate speech detection models,” in Proc. 59th Annu. Meet. Assoc. Comput. Linguistics & 11th Int. Joint Conf. NLP (Long Papers) , Online, Aug. 2021, pp. 41–58. doi:...

    2021 doi

  9. [17]

    BERT: Pre-training of deep bidirectional transformers for language understanding,

    J. Devlin, M. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidirectional transformers for language understanding,” in Proc. NAACL-HLT, Minneapolis, MN, USA, Jun. 2019, pp. 4171–4186. doi:10.18653/v1/N19-1423

    2019 doi

  10. [18]

    RoBERTa: A robustly optimized BERT pretraining approach,

    Y . Liu et al. , “RoBERTa: A robustly optimized BERT pretraining approach,” arXiv preprint arXiv:1907.11692, Jul. 2019. [Online]. Avail- able: https://arxiv.org/abs/1907.11692

    1907 arXiv

  11. [19]

    HateXplain: A benchmark dataset for explainable hate speech detection,

    B. Mathew, P. Saha, S. M. Yimam, C. Biemann, P. Goyal, and A. Mukherjee, “HateXplain: A benchmark dataset for explainable hate speech detection,” in Proc. AAAI Conf. Artif. Intell. , vol. 35, no. 17, May 2021, pp. 14867–14875. doi:10.1609/aaai.v35i17.17745

    2021 doi

  12. [20]

    Learning from the worst: Dynamically generated datasets to improve online hate detection,

    B. Vidgen, T. Thrush, Z. Waseem, and D. Kiela, “Learning from the worst: Dynamically generated datasets to improve online hate detection,” in Proc. 59th Annu. Meet. Assoc. Comput. Linguistics & 11th Int. Joint Conf. NLP (Long Papers) , Aug. 2021, pp. 1667–1682. doi:10.18653/v1...

    2021 doi

  13. [21]

    RealToxicityPrompts: Evaluating neural toxic degeneration in language models,

    S. Gehman, S. Gururangan, M. Sap, Y . Choi, and N. A. Smith, “RealToxicityPrompts: Evaluating neural toxic degeneration in language models,” in Findings Assoc. Comput. Linguistics: EMNLP 2020 , Nov. 2020, pp. 3356–3369. doi:10.18653/v1/2020.findings-emnlp.301

    2020 doi

  14. [22]

    Sarcasm detection in social media: A review,

    A. Arora, “Sarcasm detection in social media: A review,” in Proc. Int. Conf. Innov. Comput. Commun. (ICICC) , Dec. 2020, pp. 1–4. doi:10.2139/ssrn.3749018

    2020 doi

  15. [23]

    A systematic review of hate speech automatic detection using natural language processing,

    M. S. Jahan and M. Oussalah, “A systematic review of hate speech automatic detection using natural language processing,” Neurocomputing, vol. 546, Art. no. 126232, Aug. 2023. doi:10.1016/j.neucom.2023.126232

    2023

  16. [24]

    Assessing the impact of contextual information in hate speech detection,

    J. M. P ´erez et al. , “Assessing the impact of contextual information in hate speech detection,” IEEE Access , vol. 11, pp. 30575–30590, 2023. doi:10.1109/ACCESS.2023.3258973

    2023

  17. [25]

    Abusive language detection on Arabic social media,

    H. Mubarak, K. Darwish, and W. Magdy, “Abusive language detection on Arabic social media,” in Proc. 1st Workshop on Abusive Language Online, Vancouver, Canada, 2017, pp. 52–56. doi:10.18653/v1/W17- 3008

    2017 doi

  18. [26]

    Overview of the HASOC track at FIRE 2019: Hate speech and offensive content identification in Indo-European languages,

    T. Mandl et al. , “Overview of the HASOC track at FIRE 2019: Hate speech and offensive content identification in Indo-European languages,” in Proc. FIRE , Kolkata, India, 2019, pp. 14–17. doi:10.1145/3368567.3368584

    2019

  19. [27]

    'All you need is Love': Evading hate speech detection,

    T. Gr ¨ondahl, L. Pajola, M. Juuti, M. Conti, and N. Asokan, “'All you need is Love': Evading hate speech detection,” in Proc. 11th ACM Workshop Artif. Intell. Security, Toronto, Canada, 2018, pp. 2–12. doi:10.1145/3270101.3270103

    2018

  20. [28]

    Transliteration for cross-lingual morphological inflection,

    N. Murikinati, A. Anastasopoulos, and G. Neubig, “Transliteration for cross-lingual morphological inflection,” in Proc. 17th SIGMORPHON Workshop Computational Research Phonetics, Phonology, and Morphol- ogy, Online, Jul. 2020, pp. 189–197. doi:10.18653/v1/2020.sigmorphon- 1.22

    2020 doi

  21. [29]

    GLUECoS: An evaluation benchmark for code-switched NLP,

    J. Khanuja, A. Dandapat, A. Srinivasan, S. Sitaram, and M. Choud- hury, “GLUECoS: An evaluation benchmark for code-switched NLP,” in Proc. ACL-IJCNLP , Bangkok, Thailand, 2021, pp. 3575–3585. doi:10.18653/v1/2020.acl-main.329

    2021 doi

  22. [30]

    Multilingual offensive language identi- fication with cross-lingual embeddings,

    J. Ranasinghe and M. Zampieri, “Multilingual offensive language identi- fication with cross-lingual embeddings,” in Proc. EMNLP, Online, 2020, pp. 5838–5844. doi:10.18653/v1/2020.emnlp-main.470

    2020 doi

  23. [31]

    A corpus of Turkish offensive language on social media,

    C ¸ . C ¸¨oltekin, “A corpus of Turkish offensive language on social media,” in Proc. LREC , Marseille, France, 2022, pp. 4878–4885. [Online]. Available: https://aclanthology.org/2020.lrec-1.758/

    2022

  24. [32]

    Cross-domain and cross-lingual abusive language detection: A hybrid approach with deep learning and a mul- tilingual lexicon,

    E. Pamungkas and V . Patti, “Cross-domain and cross-lingual abusive language detection: A hybrid approach with deep learning and a mul- tilingual lexicon,” in Proc. ACL , Florence, Italy, 2019, pp. 363–370. doi:10.18653/v1/P19-1051

    2019 doi

  25. [33]

    LLM-Mod: Can Large Language Models Assist Content Moderation?

    Y . Liu and M. Zhang, “LLM-Mod: Can Large Language Models Assist Content Moderation?” in Proc. ACM Conf. Fairness, Accountability, and Transparency (FAccT) , Rio de Janeiro, Brazil, 2024, pp. 1–12. doi:10.1145/3613905.3650828

    2024

  26. [34]

    Respectful or toxic? Using zero-shot learning with language models to detect hate speech,

    F. M. Plaza-Del-Arco, D. Nozza, and D. Hovy, “Respectful or toxic? Using zero-shot learning with language models to detect hate speech,” in Proc. 7th Workshop Online Abuse and Harms (WOAH) , Singapore, Jan. 2023, pp. 46–52. doi:10.18653/v1/2023.woah-1.6

    2023 doi

  27. [35]

    Toxicity detection: Does context really matter?

    J. Pavlopoulos et al., “Toxicity detection: Does context really matter?” in Proc. ACL, Online, 2020, pp. 4296–4305. doi:10.18653/v1/2020.acl- main.396

    2020 doi

  28. [36]

    Just say no: Analyzing the stance of neural dialogue generation in offensive contexts,

    A. Baheti, M. Sap, and Y . Tsvetkov, “Just say no: Analyzing the stance of neural dialogue generation in offensive contexts,” in Proc. EMNLP, Online, 2021, pp. 4846–4859. doi:10.18653/v1/2021.emnlp-main.397

    2021 doi

  29. [37]

    Social bias frames: Reasoning about social and power implications of language,

    M. Sap et al. , “Social bias frames: Reasoning about social and power implications of language,” in Proc. ACL, Online, 2020, pp. 5477–5490. doi:10.18653/v1/2020.acl-main.486

    2020 doi

  30. [38]

    Process for adapting language models to society (PALMS),

    I. Solaiman and C. Dennison, “Process for adapting language models to society (PALMS),” Tech. Rep., OpenAI, 2021. [Online]. Available: https://arxiv.org/abs/2106.10328

    2021 arXiv

  31. [39]

    Man is to computer programmer as woman is to homemaker? Debiasing word embeddings,

    T. Bolukbasi, K.-W. Chang, J. Zou, V . Saligrama, and A. Kalai, “Man is to computer programmer as woman is to homemaker? Debiasing word embeddings,” in Proc. NeurIPS, Barcelona, Spain, 2016, pp. 4356–4364. doi:10.48550/arXiv.1607.06520

    2016 doi arXiv

  32. [40]

    Challenges in detoxifying lan- guage models,

    H. Welbl, A. Stiennon, and Y . Bai, “Challenges in detoxifying lan- guage models,” Tech. Rep., DeepMind, 2021. [Online]. Available: https://arxiv.org/abs/2109.07445

    2021 arXiv

  33. [41]

    Toxigen: Controllable generation of implicit and adversarial toxic text,

    R. Hartvigsen, H. Palangi, and X. He, “Toxigen: Controllable generation of implicit and adversarial toxic text,” in Proc. ACL, Dublin, Ireland, 2022, pp. 524–535. doi:10.18653/v1/2022.acl-long.39

    2022 doi

  34. [42]

    On the dangers of stochastic parrots,

    E. Bender et al., “On the dangers of stochastic parrots,” in Proc. FAccT, Online, 2021, pp. 610–623. doi:10.1145/3442188.344592

    2021

  35. [2023]

    doi:10.2147/PRBM.S397770

    doi

  36. [2024]

    Available: https://backlinko.com/social-media-users

    [Online]. Available: https://backlinko.com/social-media-users

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.