Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Lost in Pronunciation: Detecting Chinese Offensive Language Disguised by Phonetic Cloaking Replacement

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A 500-post benchmark of naturally occurring phonetically cloaked Chinese insults keeps the best tested LLM at 0.672 F1, with zero-shot chain-of-thought hurting rather than helping and Pinyin-based prompting recovering much of the lost…

desk verdict Useful real-world PCR benchmark, but the 'naturally occurring' claim is diluted by undisclosed paraphrasing; still deserves serious review. read the letter →

arxiv 2507.07640 v1 pith:JU2Q557I submitted 2025-07-10 cs.CL

Haotan Guo , Jianfei He , Jiayuan Ma , Hongbin Na , Zimu Wang , Haiyang Zhang , Qi Chen , Wei Wang
show 3 more authors
Zijing Shi Tao Shen Ling Chen
This is my paper
classification cs.CL
keywords ChineseoffensivelanguagedetectionphoneticcloakingreplacementhomophonesubstitutionPinyinpromptingcontentmoderationLLMrobustnesschain-of-thoughtRedNotedataset
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 argues that current LLM content moderation is seriously underprepared for how real Chinese users disguise insults by pronunciation rather than by the rule-based substitutions prior benchmarks used. The authors introduce phonetic cloaking replacement (PCR), a deliberate use of homophones or near-homophones to hide toxic intent, and build PCR-ToxiCN: 500 naturally occurring RedNote posts balanced between offensive and non-offensive, organized into Hanzi, Alphabet, Numerical, and Mixed replacement categories. On this dataset the best tested model reaches only 0.672 F1, false negatives dominate, and zero-shot chain-of-thought prompting pushes several models lower. The paper further shows that prompting with the toneless Pinyin transcription alongside the original text recovers much of the lost accuracy, overturning an earlier dismissal of that strategy.

What carries the argument

The load-bearing device is the toneless Pinyin prompt: the original Chinese text is transcribed into Pinyin, and the model is instructed to judge offensiveness by combining the Pinyin and the text. This puts pronunciation explicitly in front of the model, so it can hear, for instance, 'si wan' underneath '4万' before deciding whether the message is offensive. The second structural piece is the four-way surface-form taxonomy of phonetic cloaking replacement—Hanzi, Alphabet, Numerical, and Mixed replacements—which guided the curation of PCR-ToxiCN and made strategy-level diagnosis possible.

What would settle it

Collect a second, independent set of 500 naturally cloaked posts from RedNote using the same taxonomy and run the paper's standard, CoT, and Pinyin prompts without further tuning; the central claim fails if the best model clearly exceeds 0.672 F1 or if Pinyin prompting no longer beats standard prompting out of sample.

Watch

Extended reading notes

Core claim

In the paper's own terms, the discovery is that naturally occurring phonetic cloaking is a tougher and more structured threat than synthetic benchmarks suggested: a four-way taxonomy of Hanzi, Alphabet, Numerical, and Mixed replacements yields a 500-post testbed where even the strongest model, o3-mini, peaks at 0.672 F1 with 0.877 precision and 0.544 recall. Hanzi replacement is the hardest category, with most model F1 scores below 0.5, and false negatives dominate, meaning cloaked offensive posts are typically let through rather than over-blocked. Chain-of-thought prompting does not fix this; it tends to lower recall because the model reasons fluently about the literal surface text and misses the hidden pronunciation. A Pinyin-based prompt, which transcribes the text into toneless Pinyin and asks the model to judge offensiveness from both forms, lifts GPT-4o and Qwen2.5-32B close to the best thinking models and corrects the earlier judgment that this strategy is ineffective.

Load-bearing premise

The results stand or fall on whether the 500 hand-picked RedNote posts, especially the 250 labeled non-offensive, are accurate and representative ground truth; the Pinyin prompt's advantage is also measured on the same posts it was designed on.

Editorial extensions

If this is right

  • Content moderation pipelines that rely on current LLMs will systematically under-detect real Chinese phonetic evasion, especially Hanzi replacements, where most tested models fall below 0.5 F1.
  • Because false negatives dominate, the failure mode is permissive: cloaked offensive posts are likely to pass as normal text rather than being over-blocked.
  • Zero-shot chain-of-thought cannot be assumed to improve robustness in offensive-language detection; in this setting it often hurts recall and should be evaluated per task.
  • Pinyin-based prompting is a cheap, training-free mitigation that narrows the gap between smaller models and much larger reasoning models on this benchmark.
  • The taxonomy itself is a diagnostic tool: reporting scores separately for Hanzi, Alphabet, Numerical, and Mixed replacements reveals which evasion surface deserves the most defensive attention.

Reading between the lines

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

  • A test the paper does not run: an independent second wave of naturally cloaked RedNote posts would reveal whether the Pinyin prompt's gain survives outside the same 500 posts it was designed on.
  • The taxonomy's reliance on dialectal mergers such as n/l suggests that robust detection may need accent- and dialect-aware phonological modeling, a direction the paper only motivates rather than implements.
  • The same pronunciation-transcription trick could plausibly transfer to other phonographic evasions in Japanese or Korean, but the paper explicitly limits its scope to Chinese.
  • The authors' own limitations section concedes the dataset is Chinese-only and small; a larger, multi-platform collection would be needed before the strategy-level gaps, particularly Hanzi replacement, can be treated as stable facts.
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 / 5 minor

Summary. The paper introduces PCR-ToxiCN, a benchmark of 500 Chinese social-media comments containing phonetic cloaking replacements (PCR), organized into a four-way taxonomy (Hanzi, Alphabet, Numerical, and Mixed replacements). The authors evaluate seven LLMs with standard, chain-of-thought, and Pinyin-based prompting, reporting that the best model reaches only F1 0.672, that zero-shot CoT generally lowers F1, and that Pinyin-based prompting partially recovers performance on two non-thinking models. The paper also releases the dataset and full prompts in the appendix.

Significance. If the benchmark faithfully represents real user-generated PCR, the paper would provide a valuable complement to synthetic cloaking datasets such as ToxiCloakCN and would give useful evidence about the limitations of current LLMs and about CoT degradation on this task. The released dataset, the reported annotation procedure with 81.5% Fleiss kappa, the inclusion of both thinking and non-thinking models, and the full prompts in Appendix C are concrete strengths that support reproducibility. However, the paper's central 'naturally occurring' claim is undermined by the paraphrasing step described in Ethical Considerations, and the Pinyin mitigation is evaluated on only two models after error analysis on the same benchmark. These issues mean the real-world validity and the generality of the main conclusions are not yet fully established.

major comments (3)
  1. [Ethical Considerations, 'Paraphrasing' bullet; Abstract; Section 3.1] The manuscript describes PCR-ToxiCN as containing 'naturally occurring' and 'real user comments' (Abstract and Section 3.1), but Ethical Considerations states: 'Longer comments were semantically rephrased to avoid potential copyright infringement, while very short comments were retained in original form.' This means an unspecified subset of the 500 examples is researcher-edited text rather than verbatim user content. Paraphrasing can alter the exact homophonic tokens, tone, and ambiguity that define PCR, so the benchmark scores in Tables 2, 3, and 5 may reflect properties of the paraphrases rather than of actual user cloaking. The authors should quantify how many of the 500 comments were rephrased, describe the rephrasing procedure in more detail, and provide an analysis of whether paraphrased and verbatim items differ in strategy distribution, difficulty, or model error rates. Without this, the external-validity claim that these are 'naturally occurring' examples is not supported by the manuscript's own methods.
  2. [Section 4.3, Table 5] The Pinyin-based prompting strategy is tested on only GPT-4o and Qwen2.5-32B, and the design of the prompt was guided by error analysis on the same 500-instance benchmark. The reported F1 gains (0.597 to 0.651 for GPT-4o; 0.627 to 0.668 for Qwen2.5-32B) are therefore not out-of-sample estimates: the prompt was selected with access to the test answers. This raises the risk that the improvement is partly overfit to the specific errors in this benchmark. The authors should evaluate the Pinyin prompt on a held-out split or use cross-validation, and ideally extend it to more of the models in Table 2, before claiming that it 'recovers much of the lost accuracy' as a general mitigation technique.
  3. [Tables 2 and 5; Section 4.2] All quantitative claims rest on single point estimates with no confidence intervals, significance tests, or repeated runs. Several differences that support the paper's narrative are small: o3-mini (0.672) versus QwQ-32B (0.670) versus Qwen2.5-32B with Pinyin (0.668), and the CoT decline for Qwen2.5-32B (F1 0.627 to 0.545) is larger, but others are modest. Because temperature is set to 0.1 rather than 0, and the dataset has only 500 examples, sampling noise is non-negligible. The authors should report bootstrap confidence intervals, multiple seeds, or an appropriate significance test so the reader can assess whether the reported ordering of models and prompting strategies is meaningful.
minor comments (5)
  1. [Section 3.1] The phrase 'real-word dataset' appears to be a typo for 'real-world dataset'.
  2. [Section 3.1, Data Annotation] The paper reports 'an inter-annotator agreement of 81.5% for offensive labeling, as measured by Fleiss' kappa.' It would be useful to also report the raw percentage agreement and the per-category kappa for the four PCR strategies, since the strategy labels are also used in the analysis in Table 3.
  3. [Section 3.1, Data Collection] The selection process is described only as 'we follow the proposed Chinese PCR taxonomy to identify and collect' examples. More detail on the search procedure, keywords, or sampling strategy would help readers assess potential selection bias, especially because samples lacking context were removed and non-offensive examples were deliberately balanced against offensive ones.
  4. [Table 4 and Section 4.2, Effects of CoT Reasoning] The claim that CoT 'may even hinder performance' is based on a single example in Table 4. The example is illustrative, but the manuscript should make clear that the quantitative evidence is in Table 2, not in the example alone.
  5. [Appendix C, Pinyin-based Prompting] The description says the prompt combines 'Hanyu Pinyin and text below' but the exact formatting of the inserted Pinyin (e.g., toneless vs. tonal, word-segmented or character-aligned) is not specified in the prompt template. This detail is important for reproducibility and should be stated explicitly.

Circularity Check

1 steps flagged · score 2.0 of 10

No definitional circularity: the core benchmark is independent, but the Pinyin-prompt gain is measured on the same dataset that guided its selection.

  1. fitted input called prediction [Abstract; Section 4.3 (Revisiting Pinyin-based Prompting); Table 5]
    "Guided by error analysis, we revisit a Pinyin-based prompting strategy that earlier studies judged ineffective and show that it recovers much of the lost accuracy."

    The error analysis was performed on the same 500 PCR-ToxiCN posts that are then used as the evaluation set in Table 5. The Pinyin prompting strategy was therefore chosen after inspecting errors on the very benchmark on which its improvement is reported, with no held-out development set or cross-validation described. The measured gain over standard and CoT prompting is thus an in-sample selection effect, not an out-of-sample prediction, fitting the fitted-input-called-prediction pattern: the intervention is selected using the data that later certifies it. This is not a definitional tautology, but it makes the headline claim that Pinyin 'recovers much of the lost accuracy' statistically dependent on the same examples used to design it.

full rationale

PCR-ToxiCN is an empirical benchmark rather than a derivation, so the main circularity axis does not apply in its usual form. The central finding that current LLMs reach at most 0.672 F1 on the dataset is a comparison of model outputs with independent human annotations (81.5% Fleiss kappa), which is not circular. The four-way taxonomy is a descriptive grouping, and reporting per-strategy F1 differences does not reduce to a prediction made from the taxonomy itself. The only circularity-relevant step is the Pinyin prompt: it was chosen after error analysis on the same 500 examples and evaluated on the same 500 examples, so its reported benefit is an in-sample selection result rather than independent evidence. In addition, the Ethical Considerations state that 'Longer comments were semantically rephrased,' which conflicts with the abstract's 'naturally occurring' claim and weakens external validity, but this is an authenticity/correctness issue rather than a circularity. Self-citations such as Xiao et al. (2024b) are used as prior work and as a point of comparison, not as load-bearing proof of the present results. Overall score 2 reflects the one in-sample selection issue while recognizing that the benchmark itself is self-contained and externally grounded in human labels.

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

The paper introduces no free parameters fitted to data; the main quantitative choices are experimental settings (temperature 0.1, top_p 0.9, top_k 5) and the deliberate 250/250 class balance, which are design choices rather than fitted parameters. The central claims rest on four domain assumptions: annotation quality, taxonomy completeness, prompt validity, and Pinyin transcription accuracy. No new physical or theoretical entities are postulated.

assumptions (4)
  • domain assumption Human annotations of offensiveness on the 500 posts are treated as ground truth.
    All model scores are computed against these labels. Annotator agreement is 81.5% Fleiss kappa, so residual label noise enters every conclusion. Section 3.1.
  • domain assumption The four-way taxonomy (HR, AR, NR, MR) is sufficiently complete to cover real phonetic cloaking.
    Data collection followed the taxonomy, so cloaking forms outside the four categories are excluded by design and never measured. Sections 2 and 3.1.
  • domain assumption LLM responses under the supplied prompts reflect detection capability rather than prompt artifacts.
    The CoT and Pinyin comparisons vary the prompt while keeping the models fixed, so the observed differences could be specific to these prompt templates. Section 4.1.
  • domain assumption The toneless Pinyin transcription of each post is accurate and available for every sample.
    The paper states that original text was transcribed into toneless Pinyin but does not name the transcription tool or report error checks, and the method depends on that transcription being faithful. Section 4.3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Lost in Pronunciation: Detecting Chinese Offensive Language Disguised by Phonetic Cloaking Replacement." pith.science (2026). https://pith.science/paper/JU2Q557I

@misc{pith2026250707640,
  author       = {Pith},
  title        = {Pith review of: Lost in Pronunciation: Detecting Chinese Offensive Language Disguised by Phonetic Cloaking Replacement},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JU2Q557I}},
  note         = {Machine review of arXiv:2507.07640}
}
read the original abstract

Phonetic Cloaking Replacement (PCR), defined as the deliberate use of homophonic or near-homophonic variants to hide toxic intent, has become a major obstacle to Chinese content moderation. While this problem is well-recognized, existing evaluations predominantly rely on rule-based, synthetic perturbations that ignore the creativity of real users. We organize PCR into a four-way surface-form taxonomy and compile \ours, a dataset of 500 naturally occurring, phonetically cloaked offensive posts gathered from the RedNote platform. Benchmarking state-of-the-art LLMs on this dataset exposes a serious weakness: the best model reaches only an F1-score of 0.672, and zero-shot chain-of-thought prompting pushes performance even lower. Guided by error analysis, we revisit a Pinyin-based prompting strategy that earlier studies judged ineffective and show that it recovers much of the lost accuracy. This study offers the first comprehensive taxonomy of Chinese PCR, a realistic benchmark that reveals current detectors' limits, and a lightweight mitigation technique that advances research on robust toxicity detection.

Figures

Figures reproduced from arXiv: 2507.07640 by the authors.

Figure 1
Figure 1. An example of the four phonetic cloaking [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The User Interface for the Chinese Offensive Language Annotation System. It includes integrated guidelines, followed by tasks for judging offensiveness and classifying the type of phonetic variant used in the text. 10 [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗
Figure 3
Figure 3. Examples of Various Prompting Strategies Used in Experiments: Standard Prompting, Chain-of￾Thought (CoT) Prompting, and Pinyin-Based Prompting. 12 [PITH_FULL_IMAGE:figures/full_fig_p012_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 17 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Zewen Bai, Yuanyuan Sun, Shengdi Yin, Junyu Lu, Jingjie Zeng, Haohao Zhu, Liang Yang, and Hongfei Lin. 2025. https://arxiv.org/abs/2501.15451 State toxicn: A benchmark for span-level target-aware toxicity extraction in chinese hate speech detection . Preprint, arXiv:2501.15451

    arXiv 2025

  4. [4]

    Tao Cao, Hengchang Guo, Shuchen Bai, Bingbing Li, and Na Liu. 2024. https://doi.org/10.1109/ACCESS.2024.3414431 A parallel dual-channel chinese offensive language detection method combining bert and ctm topic information . IEEE Access, 12:95165--95184

    work page arXiv 2024

  5. [5]

    Eshwar Chandrasekharan, Chaitrali Gandhi, Matthew Wortley Mustelier, and Eric Gilbert. 2019. https://doi.org/10.1145/3359276 Crossmod: A cross-community learning-based system to assist reddit moderators . Proc. ACM Hum.-Comput. Interact., 3(CSCW)

    doi:10.1145/3359276 2019

  6. [6]

    Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. https://arxiv.org/abs/1703.04009 Automated hate speech detection and the problem of offensive language . Preprint, arXiv:1703.04009

    arXiv 2017

  7. [7]

    Jiawen Deng, Jingyan Zhou, Hao Sun, Chujie Zheng, Fei Mi, Helen Meng, and Minlie Huang. 2022. https://doi.org/10.18653/v1/2022.emnlp-main.796 COLD : A benchmark for C hinese offensive language detection . In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 11580--11599, Abu Dhabi, United Arab Emirates. Associat...

    doi:10.18653/v1/2022.emnlp-main.796 2022

  8. [8]

    Joseph L Fleiss. 1971. Measuring nominal scale agreement among many raters. Psychological bulletin, 76(5):378

    1971

Show all 42 references
  1. [9]

    Mirko Franco, Ombretta Gaggi, and Claudio E. Palazzi. 2023. https://doi.org/10.1145/3599696.3612895 Analyzing the use of large language models for content moderation with chatgpt examples . In Proceedings of the 3rd International Workshop on Open Challenges in Online Social Ne...

    2023

  2. [10]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Art...

    2024 arXiv

  3. [11]

    Jianfei He, Lilin Wang, Jiaying Wang, Zhenyu Liu, Hongbin Na, Zimu Wang, Wei Wang, and Qi Chen. 2024. https://doi.org/10.1109/SWC62898.2024.00246 Guardians of discourse: Evaluating llms on multilingual offensive language detection . In 2024 IEEE Smart World Congress (SWC), pag...

    2024

  4. [12]

    Chaya Hiruncharoenvate, Zhiyuan Lin, and Eric Gilbert. 2021. https://doi.org/10.1609/icwsm.v9i1.14637 Algorithmically bypassing censorship on sina weibo with nondeterministic homophone substitutions . Proceedings of the International AAAI Conference on Web and Social Media, 9(...

    2021 doi

  5. [13]

    Manoel Horta Ribeiro, Justin Cheng, and Robert West. 2023. https://doi.org/10.1145/3543507.3583275 Automated content moderation increases adherence to community guidelines . In Proceedings of the ACM Web Conference 2023, WWW '23, page 2666–2676, New York, NY, USA. Association ...

    2023

  6. [14]

    Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Mądry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, Alex Nichol, Alex Paino, and 399...

    2024 arXiv

  7. [15]

    Shagun Jhaver, Iris Birman, Eric Gilbert, and Amy Bruckman. 2019. https://doi.org/10.1145/3338243 Human-machine collaboration for content regulation: The case of reddit automoderator . ACM Trans. Comput.-Hum. Interact., 26(5)

    2019 doi

  8. [16]

    Aiqi Jiang, Xiaohan Yang, Yang Liu, and Arkaitz Zubiaga. 2022. https://doi.org/10.1016/j.osnem.2021.100182 Swsr: A chinese dataset and lexicon for online sexism detection . Online Social Networks and Media, 27:100182

    2022

  9. [17]

    Prerna Juneja, Deepika Rama Subramanian, and Tanushree Mitra. 2020. https://doi.org/10.1145/3375197 Through the looking glass: Study of transparency in reddit's moderation practices . Proc. ACM Hum.-Comput. Interact., 4(GROUP)

    2020 doi

  10. [18]

    Mahi Kolla, Siddharth Salunkhe, Eshwar Chandrasekharan, and Koustuv Saha. 2024. https://doi.org/10.1145/3613905.3650828 Llm-mod: Can large language models assist content moderation? In Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, CHI EA '24, ...

    2024

  11. [19]

    Deepak Kumar, Yousef AbuHashem, and Zakir Durumeric. 2024. https://arxiv.org/abs/2309.14517 Watch your language: Investigating content moderation with large language models . Preprint, arXiv:2309.14517

    2024 arXiv

  12. [20]

    Luan Li, Tingting Hu, and Shuting Liu. 2024. https://doi.org/10.1016/j.jml.2024.104526 Graded phonological neighborhood effects on lexical retrieval: Evidence from mandarin chinese . Journal of Memory and Language, 137:104526

    2024

  13. [21]

    Yanping Li, Catherine T Best, Michael D Tyler, and Denis Burnham. 2020. Tone variations in regionally accented mandarin. In INTERSPEECH, pages 4158--4162

    2020

  14. [22]

    Junyu Lu, Bo Xu, Xiaokun Zhang, Changrong Min, Liang Yang, and Hongfei Lin. 2023. https://doi.org/10.18653/v1/2023.acl-long.898 Facilitating fine-grained detection of C hinese toxic language: Hierarchical taxonomy, resources, and benchmarks . In Proceedings of the 61st Annual ...

    2023 doi

  15. [23]

    Huan Ma, Changqing Zhang, Huazhu Fu, Peilin Zhao, and Bingzhe Wu. 2023. Adapting large language models for content moderation: Pitfalls in data engineering and supervised fine-tuning. arXiv preprint arXiv:2310.03400

    2023 arXiv

  16. [24]

    Jiayuan Ma, Hongbin Na, Zimu Wang, Yining Hua, Yue Liu, Wei Wang, and Ling Chen. 2025 a . https://aclanthology.org/2025.coling-main.616/ Detecting conversational mental manipulation with intent-aware prompting . In Proceedings of the 31st International Conference on Computatio...

    2025

  17. [25]

    Xuchen Ma, Jianxiang Yu, Wenming Shao, Bo Pang, and Xiang Li. 2025 b . https://arxiv.org/abs/2505.22184 Breaking the cloak! unveiling chinese cloaked toxicity with homophone graph and toxic lexicon . Preprint, arXiv:2505.22184

    2025 arXiv

  18. [26]

    Todor Markov, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee, Steven Adler, Angela Jiang, and Lilian Weng. 2023. https://doi.org/10.1609/aaai.v37i12.26752 A holistic approach to undesired content detection in the real world . In Proceedings of the Thir...

    2023 doi

  19. [27]

    Mainack Mondal, Leandro Ara\' u jo Silva, and Fabr\' cio Benevenuto. 2017. https://doi.org/10.1145/3078714.3078723 A measurement study of hate speech in social media . In Proceedings of the 28th ACM Conference on Hypertext and Social Media, HT '17, page 85–94, New York, NY, US...

    2017

  20. [28]

    Sankha Subhra Mullick, Mohan Bhambhani, Suhit Sinha, Akshat Mathur, Somya Gupta, and Jidnya Shah. 2023. https://doi.org/10.18653/v1/2023.acl-industry.54 Content moderation for evolving policies using binary question answering . In Proceedings of the 61st Annual Meeting of the ...

    2023 doi

  21. [29]

    Chikashi Nobata, Joel Tetreault, Achint Thomas, Yashar Mehdad, and Yi Chang. 2016. https://doi.org/10.1145/2872427.2883062 Abusive language detection in online user content . In Proceedings of the 25th International Conference on World Wide Web, WWW '16, page 145–153, Republic...

    2016

  22. [30]

    Koustuv Saha, Eshwar Chandrasekharan, and Munmun De Choudhury. 2019. https://doi.org/10.1145/3292522.3326032 Prevalence and psychological effects of hateful speech in online college communities . In Proceedings of the 10th ACM Conference on Web Science, WebSci '19, page 255–26...

    2019

  23. [31]

    Yuqi Wang, Zimu Wang, Nijia Han, Wei Wang, Qi Chen, Haiyang Zhang, Yushan Pan, and Anh Nguyen. 2024. https://doi.org/10.18653/v1/2024.wassa-1.45 Knowledge distillation from monolingual to multilingual models for intelligent and interpretable multilingual emotion detection . In...

    2024 doi

  24. [32]

    Zimu Wang, Hongbin Na, Rena Gao, Jiayuan Ma, Yining Hua, Ling Chen, and Wei Wang. 2025. https://doi.org/10.18653/v1/2025.clpsych-1.21 From posts to timelines: Modeling mental health dynamics from social media timelines with hybrid LLM s . In Proceedings of the 10th Workshop on...

    2025 doi

  25. [33]

    Yunze Xiao, Houda Bouamor, and Wajdi Zaghouani. 2024 a . Chinese offensive language detection: Current status and future directions. arXiv preprint arXiv:2403.18314

    2024 arXiv

  26. [34]

    Yunze Xiao, Yujia Hu, Kenny Tsu Wei Choo, and Roy Ka wei Lee. 2024 b . https://arxiv.org/abs/2406.12223 Toxicloakcn: Evaluating robustness of offensive language detection in chinese with cloaking perturbations . Preprint, arXiv:2406.12223

    2024 arXiv

  27. [35]

    An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, and 23 others. 2025 a . https://arxiv.org/a...

    2025 arXiv

  28. [36]

    Shujian Yang, Shiyao Cui, Chuanrui Hu, Haicheng Wang, Tianwei Zhang, Minlie Huang, Jialiang Lu, and Han Qiu. 2025 b . https://arxiv.org/abs/2505.24341 Exploring multimodal challenges in toxic chinese detection: Taxonomy, benchmark, and findings . Preprint, arXiv:2505.24341

    2025 arXiv

  29. [37]

    i know it's sensitive

    WeiMing Ye and Luming Zhao. 2023. https://doi.org/10.1016/j.dcm.2022.100666 “i know it's sensitive”: Internet censorship, recoding, and the sensitive word culture in china . Discourse, Context & Media, 51:100666

    2023

  30. [38]

    Moira Jean Winsland Yip. 2002. Tone. Cambridge University Press

    2002

  31. [39]

    Wei Zhang and John M Levis. 2021. The southwestern mandarin/n/-/l/merger: effects on production in standard mandarin and english. Frontiers in Communication, 6:639390

    2021

  32. [40]

    Gang Zhou, Haizhou Wang, Di Jin, Wenxian Wang, Shuyu Jiang, Rui Tang, and Xingshu Chen. 2025. A toxic euphemism detection framework for online social network based on semantic contrastive learning and dual channel knowledge augmentation. Information Processing & Management, 62...

    2025

  33. [41]

    Jingyan Zhou, Jiawen Deng, Fei Mi, Yitong Li, Yasheng Wang, Minlie Huang, Xin Jiang, Qun Liu, and Helen Meng. 2022. https://doi.org/10.18653/v1/2022.findings-emnlp.262 Towards identifying social bias in dialog systems: Framework, dataset, and benchmark . In Findings of the Ass...

    2022 doi

  34. [42]

    Li Zhou, Laura Cabello, Yong Cao, and Daniel Hershcovich. 2023. https://doi.org/10.18653/v1/2023.c3nlp-1.2 Cross-cultural transfer learning for C hinese offensive language detection . In Proceedings of the First Workshop on Cross-Cultural Considerations in NLP (C3NLP), pages 8...

    2023 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.