Pith. sign in

REVIEW 4 major objections 4 minor 29 references

Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites

T0 review · 4 major / 4 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 New Chinese dataset measures whether LLMs can remove toxicity without flattening the speaker's emotion.

desk verdict The resource is real and useful; the sentiment-polarity rankings are built on a circular classifier and should not be trusted yet. read the letter →

arxiv 2505.15297 v1 pith:PWNWUF75 submitted 2025-05-21 cs.CL

Xintong Wang , Yixiao Liu , Jingheng Pan , Liang Ding , Longyue Wang , Chris Biemann This is my paper · ORCID
classification cs.CL
keywords Chinesetoxiclanguagedetoxificationsentimentpolaritytextrewritingemoji-inducedtoxicityhomophonicdialogueLLMevaluation
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 introduces ToxiRewriteCN, a Chinese-language dataset of 1,556 triplets in which every toxic sentence has a non-toxic rewrite that keeps the original sentiment polarity, plus labeled toxic spans. The dataset covers five scenarios: ordinary offensive sentences, toxicity carried by emojis, toxicity hidden in homophones, single-turn dialogues, and multi-turn dialogues. The authors use it to evaluate 17 large language models on detoxification accuracy, fluency, content preservation, and sentiment polarity. They claim that current models, including commercial and mixture-of-experts systems, can remove obvious vulgar words but frequently over-sanitize emotional tone or fail to resolve covert toxicity, especially in multi-turn conversations. The paper also reports a human preference study in which 79 of 100 annotators preferred sentiment-preserving rewrites over polite rewrites.

What carries the argument

The load-bearing object is the ToxiRewriteCN triplet: an original toxic Chinese sentence, a sentiment-aligned non-toxic rewrite, and labeled toxic spans. These triplets are produced by a six-task, human-in-the-loop pipeline: filtering by toxicity category and rewrite suitability, coarse rewriting by an LLM under instructions to preserve emotional tone while changing only toxic components, human correction, and cross-verification by a second annotator on a 5-point scale. The sentiment polarity axis is what makes the dataset novel: rewrites are required to stay in the same emotional register rather than drift into customer-service politeness. The dataset's scenario split, covering standard, emoji-induced, homophonic, single-turn, and multi-turn cases, is the mechanism that exposes where models fail.

What would settle it

Take a random sample of, say, 200 of the 1,556 toxic inputs, have independent native Chinese speakers write their own sentiment-preserving rewrites and label toxic spans, and then compare the original dataset's labels and the Qwen3-32B classifier outputs against those independent judgments; if agreement is low or the models' relative ranking changes under the new labels, the paper's central performance claims are not stable.

Watch

Extended reading notes

Core claim

The central claim is that sentiment polarity is not decoration but part of the meaning of toxic language, and that Chinese detoxification can and should be evaluated with reference rewrites that preserve it. Concretely, the paper claims that a dataset built by filtering general offensive language, rewriting with emotional guidance, and human post-editing can serve as a benchmark for this task. On this benchmark, the best closed-source generation models reach roughly 88 percent sentence-level detoxification on standard sentences, but all evaluated models degrade sharply on emoji, homophone, and multi-turn settings, where top models' sentence-level detoxification scores fall below 56 percent. The paper thus positions sentiment-preserving rewriting as a distinct capability that current LLMs have not mastered.

Load-bearing premise

The benchmark treats the human-written rewrites and the Qwen3-32B classifier labels as gold standard, but annotator agreement is not reported and the same model family that generated the initial reference rewrites is among the evaluated systems.

Editorial extensions

If this is right

  • A reusable Chinese benchmark now exists for measuring whether a model detoxifies without flattening emotion, with 1,556 reference rewrites and toxic span labels.
  • On this benchmark, closed-source and large MoE models lead overall, but no evaluated model keeps sentiment polarity consistently; reasoning models reduce polite over-sanitization at the cost of higher residual toxicity.
  • Emoji-induced and homophone-based toxicity are substantially harder than standard sentences, and multi-turn dialogue is the hardest setting, with top models' sentence-level detoxification dropping below 56 percent.
  • Human preference data indicate that neutral, sentiment-preserving rewrites are preferred to polite rewrites in 79 of 100 comparisons, supporting the task definition.
  • The main bottleneck is not producing fluent rewrites but understanding subtle toxic expressions, interpreting context, and keeping emotional tone intact.

Reading between the lines

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

  • Editorial inference: if ToxiRewriteCN is adopted as a benchmark, the emoji, homophone, and multi-turn subsets will likely become the informative axes, because standard sentences already sit near saturation for top models.
  • Editorial inference: the same polarity-preserving rewrite protocol could be adapted to code-mixed Chinese and to other high-context languages, and the sharp performance drop in multi-turn dialogues suggests discourse-level detoxification needs a separate modeling treatment.
  • Editorial inference: the observed tension between residual toxicity and over-politeness points toward a controllable emotional-intensity parameter as a natural next system component, though the paper itself does not propose one.
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 / 4 minor

Summary. The paper introduces ToxiRewriteCN, a Chinese dataset of 1,556 triplets consisting of a toxic sentence, a sentiment-aligned non-toxic rewrite, and labeled toxic spans, covering standard, emoji-induced, homophonic, single-turn, and multi-turn dialogue scenarios. The dataset is constructed through a six-step human-in-the-loop pipeline with LLM-assisted coarse rewriting followed by human post-editing and cross-verification. The authors evaluate 17 commercial and open-source LLMs on four dimensions: detoxification accuracy, fluency, content preservation, and sentiment polarity. They report that commercial and MoE models perform best overall but that all models struggle with sentiment-preserving detoxification, especially for emoji, homophone, and multi-turn dialogue inputs, where over-sanitization and under-neutralization are common.

Significance. If the dataset and evaluation are sound, ToxiRewriteCN would be a valuable resource: it is the first Chinese detoxification dataset explicitly designed to preserve sentiment polarity, and it addresses a real gap given existing Chinese toxicity datasets focus on classification rather than rewriting. The human post-editing and cross-verification pipeline is a strength, as is the decision to release the data and code. The paper also provides a broad benchmark across 17 models and fine-grained scenario analyses. However, the evaluation's reliability is weakened by the circular sentiment-polarity metric, the absence of inter-annotator agreement, and the very small scenario subsets. These issues directly bear on the paper's central claim about LLMs' inability to preserve sentiment polarity during detoxification.

major comments (4)
  1. [§4.1 and Appendix D] The Sentiment Polarity metric is circular. The three-class Qwen3-32B sentiment classifier is fine-tuned with the neutral class consisting of ToxiRewriteCN's own human-written rewrites, duplicated to a 1:2:1 ratio. Therefore, the 'Neutral↑' percentages in Tables 1–6 measure how closely model outputs resemble the gold rewrites in the eyes of a classifier trained to call those very rewrites 'neutral,' not an independent measure of emotional fidelity. The main comparative claims—for example, that GPT-4o achieves 67.16% neutral while Llama3-8B falls to 43.44%, or that dense models over-sanitize—rest on this metric. Please validate the classifier on held-out human judgments or replace it with an independently trained sentiment classifier, and report the classifier's accuracy and agreement on a held-out test set.
  2. [Tables 3, 4, and 6] The scenario-level subsets are very small (emoji n=49, homophone n=39, multi-turn n=34) and no confidence intervals or significance tests are reported. The scenario-level conclusions, such as QwQ-32B's sentence-level detoxification accuracy dropping from 85.71% overall to 24.49% on emoji inputs, or the claim that multi-turn dialogue is the hardest setting, rely on percentages that shift by several points with just a few instances. Please provide exact binomial confidence intervals or bootstrap estimates, and perform significance tests when comparing scenarios or models.
  3. [§3.4] The paper reports cross-verification with average Likert scores but provides no inter-annotator agreement statistic, such as Cohen's kappa or Krippendorff's alpha. Without a measure of agreement, the reliability of the human rewrites and toxic span labels—the gold standard for the entire benchmark—is not established. Please report agreement on a shared subset of the data.
  4. [§3.3 and Table 1] Qwen-Max generated the initial coarse rewrites, 482 of which were accepted unchanged into the final reference set, and Qwen-Max is also one of the 17 benchmarked models. This creates a contamination path: Qwen-Max's outputs are, by construction, closer to a portion of the references, which may inflate its fluency and sentiment-polarity scores. The paper should either exclude Qwen-Max from the benchmark or provide an analysis of how its scores differ on accepted versus edited reference items.
minor comments (4)
  1. [Throughout] The dataset name is rendered inconsistently as both TOXIREWRITE CN and ToxiRewriteCN; please unify the nomenclature.
  2. [§4.3] The sentence 'Reasoning models such as Deepseek-R1 and QwQ-32B reveals an interesting trade-off' contains a subject-verb agreement error; 'reveals' should be 'reveal.'
  3. [References] The reference 'X. X (formerly twitter)' contains a placeholder author name and an incomplete citation; this entry appears unfinished.
  4. [Tables 2–6] The captions state 'Box highlights' the best scores, but the boxes are not visible in the manuscript text; please ensure the final PDF displays them, or describe the best scores in the caption text.

Circularity Check

2 steps flagged · score 6.0 of 10

Sentiment polarity evaluation is self-referential: the neutral class is defined by ToxiRewriteCN's own gold rewrites, and the classifier trained on them is used to rank model outputs.

  1. self definitional [Section 4.1 (Sentiment Polarity metric) + Appendix D (Implementation Details of Classifiers)]
    "To assess the emotional tone, we apply a sentiment polarity classifier (Qwen3-32B) trained to distinguish toxic, neutral, and polite. ... The training data was constructed entirely from the TOXIREWRITE CN dataset. To improve model performance and training stability, we adjusted the label distribution to a 1:2:1 ratio of toxic, neutral, and polite examples by duplicating neutral samples, resulting in a total of 6,224 training instances."

    The 'neutral' class of the sentiment polarity classifier is literally the gold sentiment-aligned rewrites in ToxiRewriteCN; the classifier is trained entirely on that dataset and then applied to model outputs. Therefore the reported 'Neutral↑' rate is a measure of how closely a model's output resembles the gold rewrites in the eyes of a classifier trained on those very rewrites, not an independent measure of sentiment polarity. The paper's central empirical conclusion that models 'struggle to preserve sentiment polarity' (Section 4.4, Tables 1–6) rests on this metric, so the ranking can shift under an independent sentiment measure.

  2. fitted input called prediction [Section 3.3 (Task 4 and Task 5) + Section 4.2 (Models) and Table 1]
    "We use Qwen-Max, a state-of-the-art Chinese LLM, to perform initial rewrites of toxic sentences. ... Among all processed samples from Task 4, annotators accepted 482 LLM rewrites without changes, manually edited 1,085 rewrites ..."

    Qwen-Max generated the coarse rewrites that became part of the final gold set (482 accepted unchanged, the rest edited from its drafts), and Qwen-Max is one of the 17 benchmarked models in Table 1. The sentiment classifier that scores Qwen-Max's output is trained on these very gold rewrites. Hence Qwen-Max's reported sentiment polarity (64.46% neutral) and its ranking against other models are at least partly self-scored: the evaluation channel is fitted to the evaluated model's own output style. This does not determine the dataset resource itself, but it can inflate Qwen-Max's comparative scores.

full rationale

The paper's core resource, ToxiRewriteCN, is constructed through a human-in-the-loop pipeline with post-editing and cross-verification, and it is not itself circular. The main circularity is in the evaluation. The sentiment polarity metric is defined by a Qwen3-32B classifier whose neutral class is the human gold rewrites from ToxiRewriteCN; using that classifier to rank models on 'neutral' rate means the metric largely measures closeness to the dataset's own rewrites rather than an independent property of sentiment. This is load-bearing because the abstract and Section 4.4 draw their main conclusions about emotional fidelity and over-sanitization from this metric. A second, more specific channel is that Qwen-Max generated the initial rewrites that entered the gold set (482 unchanged) and is then evaluated by a classifier trained on that same gold set, giving it a structural advantage. The human preference study in Appendix C provides some independent support for preferring sentiment-preserving rewrites, but it is a small side study and does not validate the per-model neutral/polite rates in Tables 1–6. Overall, the benchmark resource is independent, but the headline comparative evaluation on sentiment polarity is partially circular, warranting a score of 6.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim relies on assumptions about annotation quality, classifier validity, and the source corpora; no new entities are introduced. Design thresholds are hand-chosen rather than fitted to a target result.

free parameters (4)
  • Cross-verification retention threshold = 4.0 average Likert score
    Samples below 4.0 were removed; chosen by hand and directly determines which 1,556 triplets survive.
  • Sentiment classifier label ratio = 1:2:1 toxic:neutral:polite
    Neutral samples duplicated to balance training; affects all Sentiment Polarity measurements.
  • Dialogue filtering caps = maximum 3 users, maximum 13 turns
    Hand-set constraints on multi-turn dialogues; shapes the conversation subsets.
  • Mild toxicity filter = only 'mild toxicity' retained
    Excludes hate speech and severe attacks by design; limits dataset to rewriteable expressions and defines scope.
assumptions (5)
  • domain assumption Source corpora (ToxiCN, COLD, ToxiCloakCN, Cdial-bias, SWSR, SCCD) provide valid toxic candidates.
    Section 3.1 samples from these datasets and trusts their toxicity annotations as the starting pool.
  • domain assumption Three native Chinese-speaking annotators' judgments are ground truth for non-toxicity, sentiment preservation, and toxic span labels.
    Section 3.4 uses their 4.0+ average ratings to retain samples; no inter-annotator agreement is reported.
  • domain assumption Fine-tuned Qwen3-32B classifiers correctly measure toxicity and sentiment polarity in model outputs.
    Section 4.1 and Appendix D rely on classifiers trained on ToxiRewriteCN without external validation on held-out or other-domain data.
  • domain assumption Text2Vec cosine similarity captures content preservation.
    Section 4.1 uses embedding cosine similarity as a proxy for semantic preservation.
  • standard math BLEU, ChrF++, BERTScore, and COMET approximate fluency.
    Standard reference-based metrics used as in prior rewriting work; their limitations are not discussed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites." pith.science (2026). https://pith.science/paper/PWNWUF75

@misc{pith2026250515297,
  author       = {Pith},
  title        = {Pith review of: Chinese Toxic Language Mitigation via Sentiment Polarity Consistent Rewrites},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PWNWUF75}},
  note         = {Machine review of arXiv:2505.15297}
}
read the original abstract

Detoxifying offensive language while preserving the speaker's original intent is a challenging yet critical goal for improving the quality of online interactions. Although large language models (LLMs) show promise in rewriting toxic content, they often default to overly polite rewrites, distorting the emotional tone and communicative intent. This problem is especially acute in Chinese, where toxicity often arises implicitly through emojis, homophones, or discourse context. We present ToxiRewriteCN, the first Chinese detoxification dataset explicitly designed to preserve sentiment polarity. The dataset comprises 1,556 carefully annotated triplets, each containing a toxic sentence, a sentiment-aligned non-toxic rewrite, and labeled toxic spans. It covers five real-world scenarios: standard expressions, emoji-induced and homophonic toxicity, as well as single-turn and multi-turn dialogues. We evaluate 17 LLMs, including commercial and open-source models with variant architectures, across four dimensions: detoxification accuracy, fluency, content preservation, and sentiment polarity. Results show that while commercial and MoE models perform best overall, all models struggle to balance safety with emotional fidelity in more subtle or context-heavy settings such as emoji, homophone, and dialogue-based inputs. We release ToxiRewriteCN to support future research on controllable, sentiment-aware detoxification for Chinese.

Figures

Figures reproduced from arXiv: 2505.15297 by the authors.

Figure 1
Figure 1. Illustration of three outcomes in detoxifying toxic Chinese sentences: (1) blocked by rule-based filters, (2) overly polite rewrites that distort user intent, and (3) sentiment-aligned detoxification that preserves emotional tone while removing toxicity. approaches often operate at the sentence level, flagging entire inputs or redacting specific tokens. However, such methods are coarse-grained and frequently over-ce… view at source ↗
Figure 2
Figure 2. Overview of the human-in-the-loop annotation pipeline. The process consists of three stages: (1) Data Filtering, where candidate toxic samples are selected and verified for rewrite suitability; (2) Rewrite with Sentiment Polarity, where LLMs perform coarse rewriting followed by human correction; and (3) Cross-verification, where annotations are validated. The output includes toxic sentences, sentiment-aligned rewrit… view at source ↗
Figure 3
Figure 3. Human post-correction interface. Annota￾tors are shown the toxic sentence and the coarse rewrite. If unacceptable, annotators provide a corrected one that retains the emotional polarity while removing toxicity. structions on rewriting goals, toxic span identifica￾tion, and sentiment polarity preservation. To ensure annotation quality and internal consis￾tency, we performed a cross-verification process. Each annotato… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Comparison of four model variants (Generation, Reasoning, MOE, and Dense) across different evaluation scenarios: overall, single-sentence, emoji, homophone, single-turn conversation, and multi-turn conversation. Each chart visualizes performance on six metrics: Detox-C…
Figure 5
Figure 5. Figure 5: Distribution of toxic data in the TOXIREWRITECN dataset. The dataset covers five distinct sources of toxicity, with direct toxic sentences and single-turn dialogues comprising the majority, while emoji-induced, homophonic, and multi-turn dialogue cases capture more nua…
Figure 7
Figure 7. Figure 7: Human preference comparison between sentiment-preserving (neutral) and over-polite detox￾ification rewrites. Annotators significantly preferred neutral rewrites (79%) over polite ones (21%). (40), and multi-turn (2). For each example, annota￾tors were shown two rewrite…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

29 extracted references · 12 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]

    Meta AI. 2025. The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation . https://ai.meta.com/blog/llama-4-multimodal-intelligence/. Accessed: 2025-05-20

    work page 2025

  4. [4]

    Mazurek, Katie Shilton, and Hal Daum \'e III

    Yang Trista Cao, Lovely-Frances Domingo, Sarah Gilbert, Michelle L. Mazurek, Katie Shilton, and Hal Daum \'e III. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.209 Toxicity detection is NOT all you need: Measuring the gaps to supporting volunteer content moderators through a user-centric method . In Proceedings of the 2024 Conference on Empirical Meth...

    doi:10.18653/v1/2024.emnlp-main.209 2024

  5. [5]

    Google DeepMind. 2025. Gemini 2.5 Pro : Our most intelligent AI model. https://blog.google/technology/google-deepmind/gemini-model-thinking-updates-march-2025/. Accessed: 2025-05-20

    work page 2025

  6. [6]

    DeepSeek-AI. 2024. https://arxiv.org/abs/2412.19437 DeepSeek-V3 technical report . Preprint, arXiv:2412.19437

    arXiv 2024

  7. [7]

    Daryna Dementieva, Nikolay Babakov, Amit Ronen, Abinew Ali Ayele, Naquee Rizwan, Florian Schneider, Xintong Wang, Seid Muhie Yimam, Daniil Moskovskiy, Elisei Stakovskii, Eran Kaufman, Ashraf Elnagar, Animesh Mukherjee, and Alexander Panchenko. 2025. https://aclanthology.org/2025.coling-main.535/ Multilingual and explainable text detoxification with parall...

    work page 2025

  8. [8]

    Daryna Dementieva, Daniil Moskovskiy, Nikolay Babakov, Abinew Ali Ayele, Naquee Rizwan, Florian Schneider, Xintong Wang, Seid Muhie Yimam, Dmitry Ustalov, Elisei Stakovskii, Alisa Smirnova, Ashraf Elnagar, Animesh Mukherjee, and Alexander Panchenko. 2024. https://ceur-ws.org/Vol-3740/paper-223.pdf Overview of the multilingual text detoxification task at P...

    work page 2024

Show all 29 references
  1. [9]

    Jiawen Deng, Jingyan Zhou, Hao Sun, Chujie Zheng, Fei Mi, Helen Meng, and Minlie Huang. 2022. https://aclanthology.org/2022.emnlp-main.796/ COLD : A benchmark for Chinese offensive language detection . arXiv preprint arXiv:2201.06025

    2022 arXiv

  2. [10]

    Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. 2024. https://cdn.openai.com/gpt-4o-system-card.pdf GPT -4o system card . arXiv preprint arXiv:2410.21276

    2024 arXiv

  3. [11]

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. 2024. https://arxiv.org/abs/2412.16720 Openai o1 system card . arXiv preprint arXiv:2412.16720

    2024 arXiv

  4. [12]

    Aiqi Jiang, Xiaohan Yang, Yang Liu, and Arkaitz Zubiaga. 2022. https://arxiv.org/abs/2108.03070 SWSR : A Chinese dataset and lexicon for online sexism detection . Online Social Networks and Media, 27:100182

    2022 arXiv

  5. [13]

    Beomseok Lee, Hyunwoo Kim, Keon Kim, and Yong Suk Choi. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.848 XD etox: Text detoxification with token-level toxicity explanations . In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages ...

    2024 doi

  6. [14]

    Varvara Logacheva, Daryna Dementieva, Sergey Ustyantsev, Daniil Moskovskiy, David Dale, Irina Krotova, Nikita Semenov, and Alexander Panchenko. 2022. https://aclanthology.org/2022.acl-long.469/ Paradetox: Detoxification with parallel data . In Proceedings of the 60th Annual Me...

    2022

  7. [15]

    Junyu Lu, Bo Xu, Xiaokun Zhang, Changrong Min, Liang Yang, and Hongfei Lin. 2023. https://aclanthology.org/2023.acl-long.898/ Facilitating fine-grained detection of Chinese toxic language: Hierarchical taxonomy, resources, and benchmarks . arXiv preprint arXiv:2305.04446

    2023 arXiv

  8. [16]

    Meta AI . 2024. https://ai.meta.com/blog/meta-llama-3 Introducing Meta LLaMA 3 : The most capable openly available LLM to date

    2024

  9. [17]

    RedNote. Rednote. https://www.xiaohongshu.com/explore. Accessed: 2025-05-20

    2025

  10. [18]

    Qwen Team. 2024. https://arxiv.org/abs/2412.15115 Qwen2.5 technical report . arXiv preprint arXiv:2412.15115

    2024 arXiv

  11. [19]

    Qwen Team. 2025. https://qwenlm.github.io/blog/qwq-32b/ QwQ-32B : Embracing the power of reinforcement learning

    2025

  12. [20]

    Xintong Wang, Jingheng Pan, Liang Ding, Longqin Jiang, Longyue Wang, Xingshan Li, and Chris Biemann. 2025. https://arxiv.org/pdf/2410.17714 Cogsteer: Cognition-inspired selective layer intervention for efficiently steering large language models . In Findings of the Association...

    2025 arXiv

  13. [21]

    Weibo. Weibo. https://www.weibo.com/. Accessed: 2025-05-20

    2025

  14. [22]

    X (formerly twitter)

    X. X (formerly twitter). https://x.com/. Accessed: 2025-05-20

    2025

  15. [23]

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

    2024 arXiv

  16. [24]

    Ming Xu. 2023. Text2vec: Text to vector toolkit. https://github.com/shibing624/text2vec

    2023

  17. [25]

    Rongwu Xu, Zian Zhou, Tianwei Zhang, Zehan Qi, Su Yao, Ke Xu, Wei Xu, and Han Qiu. 2024. https://doi.org/10.18653/v1/2024.emnlp-main.476 Walking in others' shoes: How perspective-taking guides large language models in reducing toxicity and bias . In Proceedings of the 2024 Con...

    2024 doi

  18. [26]

    Neemesh Yadav, Sarah Masud, Vikram Goyal, Md Shad Akhtar, and Tanmoy Chakraborty. 2024. https://doi.org/10.18653/v1/2024.findings-acl.831 Tox- BART : Leveraging toxicity attributes for explanation generation of implicit hate speech . In Findings of the Association for Computat...

    2024 doi

  19. [27]

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jia...

    2025 arXiv

  20. [28]

    Qingpo Yang, Yakai Chen, Zihui Xu, Yuming Shang, Sanchuan Guo, and Xi Zhang. 2025 b . https://api.semanticscholar.org/CorpusID:275821009 SCCD : A session-based dataset for Chinese cyberbullying detection . ArXiv, abs/2501.15042

    2025 arXiv

  21. [29]

    Jingyan Zhou, Jiawen Deng, Fei Mi, Yitong Li, Yasheng Wang, Minlie Huang, Xin Jiang, Qun Liu, and Helen Meng. 2022. https://arxiv.org/abs/2202.08011 Towards identifying social bias in dialog systems: Framework, dataset, and benchmark . In Findings of the Association for Comput...

    2022 arXiv

Pith tools

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