Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Breaking the Cloak! Unveiling Chinese Cloaked Toxicity with Homophone Graph and Toxic Lexicon

T0 review · 4 major / 5 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 A training-free, prompt-free method, C2TU, matches Chinese homophone-disguised words against a homo-graph and toxic lexicon, then uses BERT or LLM probabilities to filter and restore the original toxic words, reporting up to 71% higher F1…

desk verdict A clever, reproducible training-free pipeline for Chinese cloaked toxicity, undercut by an evaluation built on test cloaks generated from the method's own resources. read the letter →

arxiv 2505.22184 v2 pith:NLWK4TJV submitted 2025-05-28 cs.CL cs.AI

Xuchen Ma , Jianxiang Yu , Wenming Shao , Bo Pang , Xiang Li This is my paper
classification cs.CLcs.AI
keywords Chinesecloakedtoxicityhomophonegraphtoxiclexiconunveilingtraining-freemethodprompt-freecontentmoderationspellingcorrection
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

The paper tries to establish that deliberately cloaked toxic Chinese words—ordinary-looking homophone replacements such as 垃圾 ('rubbish') written as 辣鸡 ('spicy chicken')—can be revealed without any training or prompt engineering. Its proposed method, C2TU, first matches every substring against a hand-built homo-graph and a toxic lexicon to nominate candidate cloaked words, then uses a language model's probability estimates to keep only substitutions that make the sentence more natural, restoring the original toxic word. The authors report that this two-stage recipe outperforms the strongest competitor by up to 71% in F1 and 35% in accuracy on two public Chinese toxic datasets. A reader should care because homophonic cloaking is a cheap and common censorship-evasion trick that defeats conventional toxicity detectors, and existing uncloaking work is mostly English-specific.

What carries the argument

The load-bearing object is the Chinese homo-graph: a graph whose nodes are Chinese characters and whose edges link characters with the same tone-ignored pinyin, extended with polyphonic matches, five dialect confusions (n/l, zh/z, ch/c, sh/s, and *ng/*n), and self-loops. Together with a deduplicated toxic lexicon of proto words, it defines which substrings can be candidates. Filtering is carried by the probability-difference rule: for each $(w,l)$ pair, compute $\log(P_w/P_l)$ and replace the pair with the most negative value that is below zero, then repeat. For LLMs, Theorem 1 converts the word-probability difference into the sentence-probability difference, so the full sentence context is available despite the autoregressive left-only word probabilities.

What would settle it

Collect naturally obfuscated Chinese posts that were not produced by a homophone-substitution tool and whose gold toxic words come from a fresh annotation, then run C2TU; if sentence-level F1 falls substantially below the 74.67% reported on the constructed benchmark, the method's success is tied to the test data sharing its own homophone and lexicon resources.

Watch

Extended reading notes

Core claim

At the core is the claim that cloaked-toxicity unveiling reduces to candidate generation plus probability filtering, with no parameter updates and no prompts. On a text $X$, C2TU enumerates substrings $w$ that are homophonically aligned character-by-character with a lexicon entry $l$, producing pairs $(w,l)$. It then compares the probability of the surface form $w$ against the probability of the restored toxic form $l$ in the same context, and iteratively replaces the candidate pair with the largest log-probability gap that favors $l$. For the LLM variant, the paper proves that the word-level probability difference equals the sentence-level log-probability difference $\log P(X) - \log P(X')$, where $X'$ is the sentence with $w$ replaced by $l$; that identity is what lets an autoregressive LLM use both left and right context without changing its architecture. The BERT variant instead masks each character and reads the masked-token probabilities directly.

Load-bearing premise

The method's recall rests on the assumption that real cloaked toxic words are homophonic substitutions of words already present in the toxic lexicon and that the homo-graph contains the corresponding pronunciations; if a cloak uses a pronunciation outside the graph or a toxic word outside the lexicon, the method will never see it.

Editorial extensions

If this is right

  • A toxicity detector can run C2TU as a preprocessing normalization step, restoring cloaked words to their proto forms before classification, with no fine-tuning of the detector.
  • The same matching-and-filtering skeleton transfers to a new domain by swapping in a new toxic lexicon and homo-graph, so coverage is controlled by resources rather than by training data.
  • The LLM variant's sentence-level probability identity means uncloaking can be done with a 7B-8B model and still beat much larger generation-based baselines, reducing deployment cost.
  • The iterative replacement of the pair with the largest probability gap accounts for context changes after each correction, which the single-round ablation shows is needed for accuracy.

Reading between the lines

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

  • If C2TU is placed before an existing classifier, its benefit should grow with the cloak rate; on mildly obfuscated text the classifier may already cope, but the size of the gain is a testable prediction about the distribution of homophonic substitutions.
  • The manually corrected, deduplicated lexicon is the main bottleneck for recall, so an automatic extension would generate candidate proto words from the homo-graph's connected components rather than from a fixed word list.
  • The Bayes-based probability-difference identity could apply to other character-level text normalization tasks, such as restoring social-media abbreviations, where autoregressive LMs otherwise see only left context.
  • Because the reported cloaked test sets were constructed with the same homophone substitution tool and lexicon family, the method's real-world edge remains an open empirical question that a field deployment with naturally occurring obfuscation would answer.
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

4 major / 5 minor

Summary. The paper proposes C2TU, a training-free and prompt-free method for unveiling Chinese cloaked toxicity. It builds a homophone graph and a toxic lexicon, matches candidate toxic substrings, and then filters false positives using either BERT or an LLM. For the LLM variant, the authors prove a Bayes' theorem identity that converts word-level probability differences into sentence-level probability differences, thereby using full context despite autoregressive decoding. Experiments on ToxicloakCN and COLDataset report large gains over CSC and prompt-based baselines, with the best variant outperforming the strongest competitor by up to 71% in F1 and 35% in accuracy.

Significance. If the empirical results hold, C2TU would be a practical, parameter-free contribution to an underexplored task, and the Bayes' reformulation for LLM filtering is a clean theoretical insight that could transfer to other word-replacement tasks. The authors release code and data, and the method avoids both fine-tuning and prompt engineering. However, the significance is tempered by the evaluation's close alignment between the test cloaking process and the method's own lexicon and homo-graph; the current experiments do not demonstrate that the gains transfer to independently occurring cloaking distributions.

major comments (4)
  1. [Appendix E and Table 1] The COLDataset cloaked test set is constructed by the authors themselves: they merge the validation and test sets, apply JioNLP homophone substitution, and use the crawler keywords from COLDataset construction as the toxic lexicon. That lexicon is the same one C2TU is given (Table 9), and the homo-graph is built with tone-ignored pinyin from pypinyin, matching JioNLP's substitution rule. Likewise, ToxicloakCN's cloaks were generated by JioNLP from the ToxiCN lexicon, and C2TU's ToxicloakCN lexicon is a manually cleaned version of that same ToxiCN lexicon (Appendix E). Consequently, the matching stage is aligned with the exact noise process that produced the ground-truth corrections. The headline gains in Table 1 are therefore not established for independent cloaking distributions. Please re-evaluate on test data whose cloaking process is not derived from the method's own lexicons (e.g., human-generated cloaks or a separate substitution model), or clearly scope the claims to this aligned setting.
  2. [Table 9 and Section 4.2] The COLDataset lexicon in Table 9 contains many non-toxic crawler keywords, including province names, '女性', '男人', and other neutral terms. The F1 and accuracy scores on COLDataset therefore partly measure the restoration of crawled keywords rather than the unveiling of toxicity, and the claim that C2TU 'unveils cloaked toxicity' is conflated with general homophone matching. Please report results using a toxic-only curated lexicon, or provide a breakdown that separates toxic from non-toxic matched candidates.
  3. [Appendix E and Appendix B] Appendix E states that ToxicloakCN includes both homophone substitution and semantically similar emoji substitution, while Appendix B states that C2TU does not address emoji replacements. The main results in Table 1 do not indicate whether emoji-cloaked instances were excluded from the evaluation. If they were not excluded, the comparison is unfair to C2TU; if they were, please state the filtering criterion and the resulting dataset sizes, since this materially affects the interpretation of the reported F1 and accuracy numbers.
  4. [NeurIPS Checklist Item 7 and Table 1] The paper reports no error bars, confidence intervals, or significance tests, and the checklist acknowledges that API evaluations used single runs. Given that the main contribution is empirical and the reported gaps are large, please report variance across at least a few runs (or bootstrap estimates) for the deterministic local models, and clearly state which numbers are based on a single API call.
minor comments (5)
  1. [Figures 3 and 5] Section 4.4 and Appendix I: Figures 3 and 5 contain rendering artifacts (sequences such as '/uni00000013/uni00000014/...') and are not legible; please replace them with correctly rendered figures.
  2. [Section 4.1] The sentence 'None of the models’ parameters have been modified, nor have any prompts been used' is stated in the setup for all methods, but prompt-based baselines are used later. Please clarify that this property holds for C2TU, not for the baselines.
  3. [Algorithms 1 and 2] Algorithm 2 computes d_i as log(P_w/P_l) and then selects argmin_i d_i; since d_i is negative when P_w < P_l, the argmin selects the pair with the most negative log-ratio. Please add a brief note in the pseudocode or text to make this equivalence explicit for readers.
  4. [Abstract and Introduction] The claim of being 'the first to solve the problem under Chinese' is strong; please qualify it with respect to the exact task definition (correction of cloaked toxic words) and note that prior work [24] introduced cloaked toxicity evaluation but not a correction solution.
  5. [Section 3.1.1 and 4.1] The text refers to 'pypinyin2' in Section 3.1.1 but 'pypinyin' in Section 4.1; please unify the library name and version consistently.

Circularity Check

2 steps flagged · score 5.0 of 10

Empirical headline is partly circular: the COLDataset test cloaks are authored by the paper from the same lexicon and JioNLP homophone substitution that C2TU consumes, while the ToxicloakCN lexicon is a cleaned version of the same ToxiCN keyword list that generated its test cloaks; the Bayes-based derivation itself is independent.

  1. fitted input called prediction [Appendix E (Dataset details, COLDataset paragraph); cross-ref Table 9 and Section 3.2]
    "Since no existing work has compiled a toxic lexicon for COLDataset, we instead utilize the crawler keywords used during dataset construction[7] as toxic lexicon. Based on the lexicon, we merge the validation set with the test set and apply JioNLP [5] to perform homophone substitution and finally get the cloaked dataset, following the noise injection procedure of ToxicloakCN[24]."

    The COLDataset benchmark is a function of C2TU's own input: the same crawler-keyword list is (i) given to C2TU as its toxic lexicon and (ii) used to generate the test cloaks via JioNLP, whose tone-ignored pinyin substitution is exactly the relation encoded by C2TU's homo-graph (Section 3.1.1). The reported COLDataset F1/accuracy therefore measures restoration of cloaks generated from the method's own matching resources, not performance on an independently cloaked distribution. Many entries in Table 9 (e.g., '女性', '男人', province names) are non-toxic crawler keywords, so part of the score is keyword restoration rather than toxicity unveiling by construction.

  2. other [Appendix E (Toxic Lexicon, ToxicloakCN paragraph); cross-ref Section 4.1]
    "For keyword-based substitution, the toxic lexicon from ToxiCN is used as keywords. For cleaner lexicon without cloak, we further correct and deduplicate the ToxiCN's lexicon. We first manually correct all toxic words in the original lexicon back to their protowords, and then retain each protoword."

    ToxicloakCN's cloaks were generated by homophone-substituting ToxiCN's lexicon, and C2TU's ToxicloakCN lexicon is a manually corrected and deduplicated version of that same ToxiCN lexicon, with cloaked entries restored to protowords. The matching stage is thus pre-aligned to the exact word list that was corrupted to make the test set. The headline 'up to 71% on the F1 score and 35% on accuracy' is therefore largely a measure of how well the pipeline inverts the specific JioNLP corruption of a known keyword list, rather than evidence about arbitrary real-world Chinese cloaking, whose substitution set and lexicon are not provided by the paper.

full rationale

The core algorithmic derivation is not circular: Equation (5) is a direct Bayes' rule identity (Appendix C), and the filtering step computes probabilities from a frozen language model with no fitted parameters or labels, so no prediction is logically forced by the method's equations. There is also no load-bearing self-citation chain; the external benchmarks and public models carry the argument. However, the empirical evidence is substantially self-referential. For COLDataset, there is no pre-existing cloaked test set: the authors construct one by applying JioNLP homophone substitution to the validation+test sentences using the same crawler-keyword lexicon that C2TU is given (Appendix E, Table 9). Because the homo-graph is built with the same tone-ignored pypinyin relation that JioNLP uses, the matching stage is guaranteed to cover the exact corruption process that generates the ground truth; the remaining work is only LM filtering. For ToxicloakCN, the test cloaks were generated from the ToxiCN keyword list, and C2TU's lexicon is a manually cleaned version of that same list. The method is therefore never tested against an independently assembled cloaking distribution or a different homophone rule. Section 3.1.1 further builds the homo-graph from the characters of the evaluation dataset itself, compounding the self-reference. This does not make the internal pipeline circular, but it makes the headline performance numbers partially an evaluation of the method against its own input resources. The paper's own Appendix B concedes that emoji, character-splitting, and cross-lingual cloak types are out of scope, yet ToxicloakCN contains emoji substitutions without an explicit exclusion from Table 1, which widens the external-validity gap. Overall score 5: partial circularity in the experimental claim, with an independent derivation.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on hand-built resources: the homo-graph dialect edges, the two toxic lexicons, and the assumption that LM naturalness identifies cloaks. None of these are fitted numbers, but they are manually chosen. The test sets are generated with the same resources, so the resource choices are not independently validated.

free parameters (3)
  • Dialectal confusion edge types in homo-graph = n-l, zh-z, ch-c, sh-s, *ng-*n
    Manually added beyond tone-ignored pinyin equality (Section 3.1.1); these edges define which character substitutions count as candidates.
  • ToxicloakCN toxic lexicon = 387 words after manual correction and deduplication
    Curated by hand from the ToxiCN lexicon; used in the matching stage and not derived from first principles.
  • COLDataset toxic lexicon = 115 crawler keywords
    Taken from the keywords used to build COLDataset; used both for matching and for generating the cloaked test set.
assumptions (5)
  • domain assumption Homophonic character substitution is the dominant cloak type.
    Appendix B states other cloak types (character-splitting, cross-lingual homophones, emojis) are not addressed; the method only handles character homophones.
  • domain assumption Tone-ignored pinyin equality plus five dialectal confusions captures real cloaking.
    Section 3.1.1 defines these homo-graph edges without empirical evidence that they match real cloak distributions.
  • domain assumption The curated toxic lexicon covers the cloaks present in the test data.
    Test sets are generated from the same lexicon family (Appendix E), so coverage is assumed rather than validated on independent data.
  • domain assumption LM probability is a valid naturalness score for judging whether a replacement is a cloak.
    The filtering stage assumes the proto toxic word yields higher LM probability than the cloaked word in context.
  • standard math Bayes' rule applies to auto-regressive sequence probabilities.
    Theorem 1 and Appendix C use P(w|Xpre,Xtail)=P(X)/P(Xpre,Xtail); standard probability.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Breaking the Cloak! Unveiling Chinese Cloaked Toxicity with Homophone Graph and Toxic Lexicon." pith.science (2026). https://pith.science/paper/NLWK4TJV

@misc{pith2026250522184,
  author       = {Pith},
  title        = {Pith review of: Breaking the Cloak! Unveiling Chinese Cloaked Toxicity with Homophone Graph and Toxic Lexicon},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NLWK4TJV}},
  note         = {Machine review of arXiv:2505.22184}
}
abstract

Social media platforms have experienced a significant rise in toxic content, including abusive language and discriminatory remarks, presenting growing challenges for content moderation. Some users evade censorship by deliberately disguising toxic words through homophonic cloak, which necessitates the task of unveiling cloaked toxicity. Existing methods are mostly designed for English texts, while Chinese cloaked toxicity unveiling has not been solved yet. To tackle the issue, we propose C$^2$TU, a novel training-free and prompt-free method for Chinese cloaked toxic content unveiling. It first employs substring matching to identify candidate toxic words based on Chinese homo-graph and toxic lexicon. Then it filters those candidates that are non-toxic and corrects cloaks to be their corresponding toxicities. Specifically, we develop two model variants for filtering, which are based on BERT and LLMs, respectively. For LLMs, we address the auto-regressive limitation in computing word occurrence probability and utilize the full semantic contexts of a text sequence to reveal cloaked toxic words. Extensive experiments demonstrate that C$^2$TU can achieve superior performance on two Chinese toxic datasets. In particular, our method outperforms the best competitor by up to 71% on the F1 score and 35% on accuracy, respectively. Our code and data are available at https://github.com/XDxc-cuber/C2TU-Chinese-cloaked-toxicity-unveiling.

Figures

Figures reproduced from arXiv: 2505.22184 by the authors.

Figure 1
Figure 1. The main workflow of C2TU method. C2TU consists of two key stages: matching and filtering. In the matching stage, we identify candidate toxic terms in the input text by leveraging a homo-graph and a toxic lexicon. In the filtering stage, we use a language model (BERT model or LLM) to compute the probability gap for each (w, l) pair and iteratively replace the pair with the most significant gap. Once the iteration te… view at source ↗
Figure 2
Figure 2. Three candidate toxic words pairs are matched in the input text. Each pair is individually [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Efficiency study on the sentence-level correction task. We next analyze the model efficiency. We conduct experiments on two datasets, where ▲ indicates COLDataset and ⋆ represents ToxicloakCN. We use different colors to de￾note different methods including C2TU-LLM, C 2TU-BERT, Simple-CSC and Baichuan2-7B￾Base. For fairness, all LLM-based methods are based on Baichuan2-7B-Base. The results are presented in [PITH_FUL… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Sentence level correction accuracy gap between C [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Full results of ablation study. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

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

    cs.CL 2025-07 conditional novelty 6.0 of 10

    A new 500-post benchmark of naturally occurring phonetic cloaking shows LLMs detect such Chinese offensive language with F1 at most 0.672, and Pinyin-augmented prompting partially repairs the gap.

Reference graph

Works this paper leans on

33 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [1]

    Llama 3 model card

    AI@Meta. Llama 3 model card. 2024

    2024

  2. [2]

    Arora, P

    A. Arora, P. Nakov, M. Hardalov, S. M. Sarwar, V . Nayak, Y . Dinkov, D. Zlatkova, K. Dent, A. Bhatawdekar, G. Bouchard, et al. Detecting harmful content on online platforms: what platforms need vs. where research efforts go.ACM Computing Surveys, 56(3):1–17, 2023

    work page 2023

  3. [3]

    T. Bayes. An essay towards solving a problem in the doctrine of chances.Biometrika, 45(3- 4):296–315, 1958

    work page 1958

  4. [4]

    Cheng, W

    X. Cheng, W. Xu, K. Chen, S. Jiang, F. Wang, T. Wang, W. Chu, and Y . Qi. Spellgcn: Incorporating phonological and visual similarities into language models for chinese spelling check.arXiv preprint arXiv:2004.14166, 2020

    arXiv 2004

  5. [5]

    C. Cui. Jionlp, 2020

    work page 2020

  6. [6]

    Davidson, D

    T. Davidson, D. Warmsley, M. Macy, and I. Weber. Automated hate speech detection and the problem of offensive language. InProceedings of the international AAAI conference on web and social media, volume 11, pages 512–515, 2017

    2017

  7. [7]

    J. Deng, J. Zhou, H. Sun, C. Zheng, F. Mi, H. Meng, and M. Huang. Cold: A benchmark for chinese offensive language detection.arXiv preprint arXiv:2201.06025, 2022

    arXiv 2022

  8. [8]

    Devlin, M.-W

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technolo- gies, volume 1 (long and short papers), pages 4171–4186, 2019

    2019

Show all 33 references
  1. [9]

    G. Dong, J. Zhao, T. Hui, D. Guo, W. Wang, B. Feng, Y . Qiu, Z. Gongque, K. He, Z. Wang, et al. Revisit input perturbation problems for llms: A unified robustness evaluation framework for noisy slot filling task. InCCF International Conference on Natural Language Processing an...

    2023

  2. [10]

    M. Dong, Y . Chen, M. Zhang, H. Sun, and T. He. Rich semantic knowledge enhanced large language models for few-shot chinese spell checking.arXiv preprint arXiv:2403.08492, 2024

    2024 arXiv

  3. [11]

    Ghosh, P

    S. Ghosh, P. Varshney, E. Galinkin, and C. Parisien. Aegis: Online adaptive ai content safety moderation with ensemble of llm experts.arXiv preprint arXiv:2404.05993, 2024

    2024 arXiv

  4. [12]

    H. Gong, Y . Li, S. Bhat, and P. Viswanath. Context-sensitive malicious spelling error correction. InThe World Wide Web Conference, pages 2771–2777, 2019

    2019

  5. [13]

    Jafari, J

    N. Jafari, J. Allan, and S. M. Sarwar. Target span detection for implicit harmful content. InProceedings of the 2024 ACM SIGIR International Conference on Theory of Information Retrieval, pages 117–122, 2024

    2024

  6. [14]

    Jiang, X

    A. Jiang, X. Yang, Y . Liu, and A. Zubiaga. Swsr: A chinese dataset and lexicon for online sexism detection.Online Social Networks and Media, 27:100182, 2022

    2022

  7. [15]

    H. Kirk, B. Vidgen, P. Rottger, T. Thrush, and S. A. Hale. Hatemoji: A test suite and adversarially-generated dataset for benchmarking and detecting emoji-based hate. In M. Carpuat, M.-C. de Marneffe, and I. V . Meza Ruiz, editors,Proceedings of the 2022 Conference of the Nort...

    2022

  8. [16]

    J. Li, Q. Wang, Z. Mao, J. Guo, Y . Yang, and Y . Zhang. Improving chinese spelling check by character pronunciation prediction: The effects of adaptivity and granularity.arXiv preprint arXiv:2210.10996, 2022

    2022 arXiv

  9. [17]

    K. Li, Y . Hu, L. He, F. Meng, and J. Zhou. C-llm: Learn to check chinese spelling errors character by character.arXiv preprint arXiv:2406.16536, 2024. 11

    2024 arXiv

  10. [18]

    A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437, 2024

    2024 arXiv

  11. [19]

    J. Lu, B. Xu, X. Zhang, C. Min, L. Yang, and H. Lin. Facilitating fine-grained detection of chinese toxic language: Hierarchical taxonomy, resources, and benchmarks.arXiv preprint arXiv:2305.04446, 2023

    2023 arXiv

  12. [20]

    Phute, A

    M. Phute, A. Helbling, M. Hull, S. Peng, S. Szyller, C. Cornelius, and D. H. Chau. Llm self de- fense: By self examination, llms know they are being tricked.arXiv preprint arXiv:2308.07308, 2023

    2023 arXiv

  13. [21]

    X. Rao, Y . Zhang, S. Peng, Q. Jia, and X. Liu. Chinese hate speech detection method based on RoBERTa-WWM. In M. Sun, B. Qin, X. Qiu, J. Jiang, and X. Han, editors,Proceedings of the 22nd Chinese National Conference on Computational Linguistics, pages 501–511, Harbin, China, A...

    2023

  14. [22]

    Sheth, V

    A. Sheth, V . L. Shalin, and U. Kursuncu. Defining and detecting toxicity on social media: context and knowledge are key.Neurocomputing, 490:312–318, 2022

    2022

  15. [23]

    Q. Team. Qwen2.5: A party of foundation models, September 2024

    2024

  16. [24]

    Y . Xiao, Y . Hu, K. T. W. Choo, and R. K.-w. Lee. Toxicloakcn: Evaluating robustness of offen- sive language detection in chinese with cloaking perturbations.arXiv preprint arXiv:2406.12223, 2024

    2024 arXiv

  17. [25]

    A. Yang, B. Xiao, B. Wang, B. Zhang, C. Bian, C. Yin, C. Lv, D. Pan, D. Wang, D. Yan, et al. Baichuan 2: Open large-scale language models.arXiv preprint arXiv:2309.10305, 2023

    2023 arXiv

  18. [26]

    Zhang, H

    H. Zhang, H. Gao, Q. Hu, G. Chen, L. Yang, B. Jing, H. Wei, B. Wang, H. Bai, and L. Yang. Chinesesafe: A chinese benchmark for evaluating safety in large language models.arXiv preprint arXiv:2410.18491, 2024

    2024 arXiv

  19. [27]

    Zhang, H

    S. Zhang, H. Huang, J. Liu, and H. Li. Spelling error correction with soft-masked bert.arXiv preprint arXiv:2005.07421, 2020

    2005 arXiv

  20. [28]

    Y . Zhao, J. Zhu, C. Xu, and X. Li. Enhancing llm-based hatred and toxicity detection with meta-toxic knowledge graph.arXiv preprint arXiv:2412.15268, 2024

    2024 arXiv

  21. [29]

    H. Zhou, Z. Li, B. Zhang, C. Li, S. Lai, J. Zhang, F. Huang, and M. Zhang. A simple yet effective training-free prompt-free approach to chinese spelling correction based on large language models.arXiv preprint arXiv:2410.04027, 2024

    arXiv

  22. [30]

    J. Zhou, J. Deng, F. Mi, Y . Li, Y . Wang, M. Huang, X. Jiang, Q. Liu, and H. Meng. Towards identifying social bias in dialog systems: Framework, dataset, and benchmark. In Y . Goldberg, Z. Kozareva, and Y . Zhang, editors,Findings of the Association for Computational Linguist...

    2022

  23. [31]

    Input-output length consistency is mandatory

  24. [32]

    魄”, “蠢”, “鸡

    Return only the corrected sentence without any additional output. - User: <Input sentence> - Assistant: <Output sentence> H Toxic Lexicon We present toxic lexicons of ToxicloakCN and COLDataset here, splitting each lexicon by toxic words’ length. Specifically, the toxic lexico...

  25. [33]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

    2025

Pith tools

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