Pith. sign in

REVIEW 4 major objections 4 minor 34 references

Span-level rationales shift the toxicity–meaning trade-off in multilingual detoxification; they do not guarantee better results.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 17:13 UTC pith:DS7TOZXJ

load-bearing objection An honest, modest exploratory study of span-guided detoxification; the conditional-usefulness finding is plausible, but the toxicity numbers need an external scorer before the central trade-off claim is fully supported. the 4 major comments →

arxiv 2607.17713 v1 pith:DS7TOZXJ submitted 2026-07-20 cs.CL

AEGIS: Awareness-Enhanced Guidance for Iterative Safeguard

Kyungwon Park , Sangmin Lee , Heejae Chon , Hyungu Kang This is my paper
classification cs.CL
keywords span-level rationalestext detoxificationmultilingual detoxificationcontrollabilitytoxicity–meaning trade-offrationale-guided generationself-reflectionfrozen generator
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper asks when span-level rationales help or hurt text detoxification, and answers: conditionally. AEGIS separates a fixed span detector from a frozen generator, feeding harmful spans, intensity labels, and target attributes into the rewrite prompt, with optional self-reflection. Testing across English, Chinese, and Korean and across two generator families, the authors find that guidance consistently moves the balance between toxicity reduction and meaning preservation, but neither guided rewriting nor reflection uniformly beats unguided baselines. The paper's central claim is that explicit rationales are a controllability mechanism—a dial to tune where a rewrite lands on the toxicity–meaning frontier—rather than a guarantee of superior detoxification.

Core claim

On the paper's own terms, the central discovery is that span-guided detoxification is conditionally useful. In the key ablation, adding span guidance to a frozen generator raises BERTScore from 0.82 to 0.90 and LLM-judged semantic similarity from 1.87 to 3.00 relative to unguided rewriting, while adding self-reflection cuts toxicity from 0.35 to 0.25 at a small BERTScore cost (0.90 to 0.88). Across languages and backbones, no single configuration dominates: LLaMA 3.2 variants preserve meaning better in English, while Qwen 3 variants obtain stronger LLM-based toxicity-reduction scores in some Chinese and Korean settings. The paper therefore claims that the effect of explicit rationales depend

What carries the argument

The load-bearing mechanism is AEGIS's modular split: a fixed span-level detector produces five-way intensity-aware BIO tags (SOFT/HARD spans), target-attribute labels, and sentence-level toxicity classes; these are turned into natural-language prompts that instruct the generator to paraphrase low-intensity spans gently and delete or replace high-intensity spans strongly. An optional self-reflection loop reuses the same detector as a critic, gating up to three refinement iterations on toxicity and fidelity thresholds. Because the detector is fixed and generators are frozen, any observed differences across backbones must come from how each generator responds to identical span-level guidance.

Load-bearing premise

The fixed detector is a valid instrument both for producing the spans that guide rewriting and for measuring the toxicity of the output; if the in-house detector dominates the reported toxicity numbers, part of the observed reduction is the generator satisfying the same detector that wrote the guidance, which may not transfer to real-world harm reduction.

What would settle it

Re-run the guided versus unguided conditions using an independent, pre-existing toxicity scorer (for example, Perspective API) as the sole outcome measure, without any local detector in the loop; if the guided advantage in toxicity reduction shrinks or disappears while BERTScore differences stay, the claim that explicit rationales shift the trade-off would be weakened.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the conditional-usefulness claim is right, detoxification pipelines should be tuned per generator and per language; a single guided-prompt recipe will not win across backbones.
  • Guidance and reflection act as orthogonal controls: practitioners can trade a small meaning-preservation cost for lower toxicity by adding reflection, or prioritize semantic fidelity by stopping after the guided rewrite.
  • The result warns against generalizing monolingual or single-backbone studies of rationale-guided rewriting; conclusions from English-only or LLaMA-only experiments may not transfer.
  • The fixed-detector/frozen-generator design offers a reusable instrumentation for studying controllability without fine-tuning, which is useful for multilingual deployment where labeled data is scarce.
  • If the detector's own spans are used as the outcome measure, gains may partly reflect the generator being coaxed to satisfy the same critic that wrote the rationales—so independent scoring is essential.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The conditional-usefulness result suggests a broader principle: any explicit rationale signal acts as a prior that a generator may overfit; whether it helps depends on the model's capacity to exploit the signal, so guidance may help small models more than large ones.
  • A testable extension is to compare the same guided prompt across a wider range of generator scales: if very strong LLMs already avoid toxicity, the guidance effect should shrink, which would confirm the claim's conditionality.
  • The framework could be extended to other control tasks—e.g., style transfer or named-entity masking—where the same question 'does explicit span grounding help or hurt?' may have similarly conditional answers.
  • Because the paper uses only automatic and LLM-based metrics, a human-judgment study would be the decisive next step to verify that the toxicity–meaning trade-off it measures reflects real-world perception.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper presents AEGIS, a modular framework for multilingual span-guided detoxification. A fixed span-level detector supplies rationales (harmful spans, intensity, target attributes) to frozen generator LLMs (LLaMA 3.2 and Qwen 3) across English, Mandarin Chinese, and Korean. The paper explicitly disclaims state-of-the-art performance and instead studies when span guidance helps or hurts the toxicity–meaning trade-off. The main evidence is Table 7, which shows that adding span guidance improves BERTScore (0.82→0.90) and LLM semantic-similarity (1.87→3.00) over unguided generation, while reflection further reduces toxicity (0.35→0.25) at a small BERTScore cost (0.90→0.88). Table 6 reports backbone- and language-dependent patterns. The paper concludes that span-guided detoxification is conditionally useful and should be viewed as a controllability dial rather than a uniformly superior method.

Significance. If the central claim holds, the paper provides a useful diagnostic contribution: it reframes span rationales as a controllability mechanism that shifts the toxicity–meaning frontier rather than a free improvement. Strengths include the modular separation of detector and frozen generators, the multilingual span-annotation setup, the explicit exploratory framing, detector validation as instrumentation (Table 8), and qualitative examples across three languages. However, the current evidentiary base is fragile in ways that the paper itself does not fully acknowledge. The most serious issue is evaluator overlap: the same detector emits the span guidance, acts as the reflection critic, and may be one of the scorers behind the reported toxicity reductions. The paper also omits key experimental details for the headline ablation (Table 7) and reports no variance or sample sizes. These issues make the central quantitative trade-off claim conditional on external validation, although the qualitative conclusion likely survives.

major comments (4)
  1. [§4.3 / Tables 6–7] The paper states that toxicity is computed using 'both the Perspective API and our detector's span-level scores,' but Tables 6–7 never identify which scorer produced the reported Tox.↓ values. Since the same detector supplies the span guidance (§3.1) and acts as the reflection critic (§3.2), using it as the toxicity evaluator creates a circularity: the reported guided/reflection toxicity reductions (e.g., 0.84→0.35→0.25 in Table 7) may partly reflect the generator being iteratively revised until the detector that wrote the spans accepts the output. The divergence with the external LLM-T scores (unguided 4.37 vs guided 4.00 while Tox.↓ improves) is consistent with this confound. Please report Perspective API and detector scores separately for all conditions, and re-score with a fixed external detector not used anywhere in the loop.
  2. [§5.2 / Table 7] This ablation is the primary evidence for the conditional-usefulness claim, but the table does not identify the language, generator backbone, model size, sample size, decoding parameters, or the reflection thresholds (θtoxicity, σfidelity) used. Without these details the result cannot be reproduced or compared across conditions, and the reader cannot tell whether the guidance/reflection effects are driven by a specific configuration. Please provide per-language and per-backbone versions, or state clearly that the table aggregates across them and report the breakdown.
  3. [§3.2] The self-reflection loop is gated by thresholds θtoxicity and σfidelity and runs for up to three iterations, but the threshold values and the number of iterations actually used are never given. Because the critic is the same detector that produced the guidance, the reflection behavior is likely detector-specific. Report the threshold values, the distribution of iterations used across examples, and a sensitivity analysis around θtoxicity to show that the iteration-wise trade-off in Figure 3 is not an artifact of a single arbitrarily chosen setting.
  4. [Tables 6–7] No variance, confidence intervals, or sample sizes are reported for any evaluation number. The observed differences (e.g., BERTScore 0.90 vs 0.88, Tox.↓ 0.35 vs 0.25) may be within sampling noise; without per-condition counts and dispersion measures, the trade-off interpretation is not quantitatively supported. For an exploratory study, at least report n per condition and standard deviations or bootstrap intervals.
minor comments (4)
  1. [§5.2 / Figure 3] Figure 3 is referenced in the text but the plots are not included in the manuscript body. Please add the figure or adjust the reference.
  2. [Table 2 caption] The prompt template has inconsistent capitalization after the colon: 'For lower-intensity spans (SOFT), paraphrase gently; For higher-intensity spans (HARD), delete or replace more strongly.' Use lower-case 'for' after the semicolon.
  3. [Table 6] The caption states that bold indicates the best and underline the second-best value within each language block, but the table formatting in the submitted text does not show these highlights. Please ensure the final PDF displays them correctly.
  4. [§5.3 / Table 8] Table 8 reports only F1 scores. For span-extraction instrumentation, also report precision and recall—especially Non-O F1—since the gap between precision and recall may affect the quality of the extracted rationales and downstream guidance.

Circularity Check

2 steps flagged

Partial circularity: the detector that supplies span guidance also gates reflection and may be the reported Tox.↓ scorer; Figure 3(d) improves by construction.

specific steps
  1. self definitional [§5.2 (Impact of Iterative Guidance), Figure 3(d); §3.2]
    "(d) Detector-safe rate improvement over iterations. ... This process is repeated for up to three iterations or until all metrics are satisfied."

    The detector-safe rate is the fraction of outputs accepted by the detector, and the algorithm explicitly loops until the detector's metrics are satisfied. The plotted 'improvement over iterations' is therefore guaranteed by the stopping rule, not an empirical effect of iterative guidance. Presenting it as an outcome of reflection makes a self-definitional quantity look like an independent finding.

  2. other [§3.2; §4.3; Tables 6–7]
    "it is evaluated by the detector module (we call the critic model), which compares the generated text against the original sentence using automatic metrics such as BERTScore and toxicity. ... In our experiments, we use both the Perspective API and our detector’s span-level scores to calculate toxicity."

    The same detector emits the span rationales used as guidance and acts as the critic whose thresholds terminate reflection, while §4.3 names the detector's span-level scores as one of the two toxicity measures. Tables 6–7 never state which scorer produced Tox.↓. If the reported Tox.↓ is detector-based, the Guided/Reflect toxicity numbers are partly a measure of whether the output satisfies the detector that wrote the guidance, not an independent outcome. External LLM-T scores in Table 7 do not show the same guided advantage (unguided 4.37 vs guided 4.00), consistent with this overlap, and the threshold values θtoxicity, σfidelity are never given.

full rationale

The paper's central qualitative claim—that span guidance changes the toxicity–meaning trade-off and is conditional on generator and language—is supported partly by independent signals: BERTScore, LLM-S, and LLM-T from a Gemini 2.5 Pro judge. Those external judgments keep the main claim from being wholly circular. However, two parts of the derivation chain fold back onto the paper's own instrument. First, Figure 3(d) reports 'detector-safe rate improvement over iterations' as if it were an empirical benefit of reflection, but §3.2 makes the stopping rule 'repeat... until all metrics are satisfied' by the detector, so the improvement is definitional. Second, §3.2 and §4.3 leave open the possibility that the reported Tox.↓ is the same in-house detector that generated the spans and gated the reflection loop; if so, the toxicity reductions in Tables 6–7 partly measure agreement with the pipeline's own detector rather than external harm reduction. The divergence between Tox.↓ and LLM-T in Table 7 is consistent with that concern. The paper also never reports the threshold values used for reflection, which prevents auditing the loop. These issues make the quantitative toxicity results partially circular, but because the central conditional-usefulness conclusion does not rest solely on the detector-based metric, a score of 4 is appropriate rather than a higher score.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 0 invented entities

The central analysis is an empirical comparison, so it rests on five domain assumptions about data quality, intensity-label provenance, instrument comparability, and metric validity — plus four fitted or ad hoc parameters (θspan, β, per-language class weights, reflection thresholds, the last unnamed). No new entities are postulated: AEGIS's components (multilingual encoders, CRF decoding, frozen-LLM prompting, self-reflection) are all prior techniques; the 5-way SOFT/HARD BIO scheme is an ordinal labeling choice whose intensity provenance is partly unexplained rather than an independently falsifiable entity.

free parameters (4)
  • θspan (span confidence threshold) = 0.6
    Filters unreliable detected spans before they become generator guidance; 'determined on the validation set' (§3.1, Output Span Integration).
  • β (BIO loss weight) = 2.0
    Chosen by hand to mitigate the <1% non-O token class imbalance (§4.2, Loss Functions).
  • Per-language BIO class weights = EN: 10/10/20/20; KO: 8/4/15/8; ZH: 8/6/15/12 (SOFT/HARD)
    Hand-set weights for B/I-SOFT and B/I-HARD (Table 4); presuppose intensity labels that §4.1 does not document for English or Chinese.
  • Reflection thresholds θtoxicity, σfidelity and iteration cap = not reported (cap: 3 iterations)
    Gate the self-reflection loop (§3.2); without numeric values the refinement process cannot be replicated and its effect cannot be independently audited.
axioms (5)
  • domain assumption HateXplain, STATE ToxiCN, and K-HATERS provide accurate, comparable span/intensity/target annotations across three languages.
    All of §5's trade-off claims presuppose the detector learns real harmful spans from these corpora; HateXplain's IAA is only 0.61 (§4.1) yet its labels are treated as ground truth for the fixed instrument.
  • domain assumption SOFT/HARD intensity is well-defined and derivable in all three languages.
    Table 4 applies intensity class weights in English and Chinese, but §4.1 describes intensity grades (L1/L2) only for K-HATERS; the derivation of English/Chinese intensity labels is never stated (§3.1, Table 4).
  • domain assumption The detector is a fixed, language-comparable instrument, so cross-generator and cross-language differences are attributable to the generator.
    §3 states the detector is 'fixed when comparing frozen generator backbones', yet §5.3 reports detector F1 differences across languages (e.g., Korean BIO F1 97 vs Chinese 95) that are then ignored in the cross-lingual comparisons.
  • domain assumption Automatic metrics and Gemini 2.5 Pro judgments are valid proxies for toxicity reduction and meaning preservation.
    §4.3 relies on BERTScore, Perspective API, in-house detector scores, and an unvalidated LLM judge; no human-evaluation anchor is provided (acknowledged in Appendix A).
  • domain assumption BERTScore is comparable across languages with different tokenizers and pretrained embedding models.
    Table 6 reports BERTScore across EN/ZH/KO; BERTScore's language-specific embedding backbones make cross-lingual numeric comparisons of questionable validity, and the paper does not address this.

pith-pipeline@v1.3.0-alltime-deepseek · 9929 in / 22935 out tokens · 227396 ms · 2026-08-01T17:13:12.614357+00:00 · methodology

0 comments
read the original abstract

Span-level rationales are often assumed to improve controllability in text detoxification, but it remains unclear when such guidance helps and when it introduces trade-offs. We present Awareness-Enhanced Guidance for Iterative Safeguard (AEGIS) as an exploratory framework for studying span-guided multilingual detoxification across English, Mandarin Chinese, and Korean. AEGIS combines span-level detector outputs with frozen generator backbones, allowing harmful spans, intensity labels, and target attributes to be provided as structured guidance during rewriting. Rather than claiming state-of-the-art detoxification performance, we analyze how span guidance affects the balance between toxicity reduction and meaning preservation across generator families, model scales, and languages. Our results suggest that span-guided detoxification is conditionally useful: explicit rationales change the trade-off between toxicity reduction and meaning preservation, but their effects depend strongly on the generator backbone and the linguistic context. These findings highlight both the promise and the limitations of span-level control signals for multilingual detoxification.

Figures

Figures reproduced from arXiv: 2607.17713 by Heejae Chon, Hyungu Kang, Kyungwon Park, Sangmin Lee.

Figure 1
Figure 1. Figure 1: Overall framework of AEGIS. ator interprets the guidance too aggressively. This paper therefore does not aim to establish a new state-of-the-art detoxification system. Instead, we ask a more diagnostic question: when does span￾level guidance help multilingual detoxification, and when does it hurt? We study this question using Awareness￾Enhanced Guidance for Iterative Safeguard (AEGIS), a modular framework … view at source ↗
Figure 2
Figure 2. Figure 2: Specific architecture and flow diagram of the detector and generator. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Iteration-wise behavior of AEGIS under self-refinement. The plots illustrate toxicity–meaning trade-offs [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

34 extracted references · 9 canonical work pages · 1 internal anchor

  1. [1]

    K-HATERS: A Hate Speech Detection Corpus in Korean with Target-Specific Ratings

    Park, Chaewon and Kim, Soohwan and Park, Kyubyong and Park, Kunwoo. K-HATERS: A Hate Speech Detection Corpus in Korean with Target-Specific Ratings. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023. doi:10.18653/v1/2023.findings-emnlp.952

    work page doi:10.18653/v1/2023.findings-emnlp.952 2023

  2. [2]

    Proceedings of the 29th International Conference on Computational Linguistics (COLING 2022) , pages =

    Why Is It Hate Speech? Masked Rationale Prediction for Explainable Hate Speech Detection , author =. Proceedings of the 29th International Conference on Computational Linguistics (COLING 2022) , pages =. 2022 , address =

    2022

  3. [3]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  4. [4]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  5. [5]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  6. [6]

    Dan Gusfield , title =. 1997

    1997

  7. [7]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  8. [8]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  9. [9]

    Neurocomputing , volume =

    A Systematic Review of Hate Speech Automatic Detection Using Natural Language Processing , author =. Neurocomputing , volume =. 2023 , publisher =. doi:10.1016/j.neucom.2023.126232 , url =

    arXiv 2023

  10. [10]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =

    Proactively Reducing the Hate Intensity of Online Posts via Hate Speech Normalization , author =. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages =. 2022 , organization =. doi:10.1145/3534678.3539161 , url =

    arXiv 2022

  11. [11]

    arXiv preprint arXiv:2403.19836 , year=

    Target Span Detection for Implicit Harmful Content , author=. arXiv preprint arXiv:2403.19836 , year=

    Pith/arXiv arXiv

  12. [12]

    Findings of the Association for Computational Linguistics: EACL 2023 , pages =

    EasyMix and Entailment Recasting for Hate Speech Detection , author =. Findings of the Association for Computational Linguistics: EACL 2023 , pages =. 2023 , url =

    2023

  13. [13]

    arXiv preprint arXiv:2307.05578 , year =

    Hate Speech Detection via Dual Contrastive Learning , author =. arXiv preprint arXiv:2307.05578 , year =

    Pith/arXiv arXiv

  14. [14]

    Scientific Reports , volume =

    Combating hate speech using an adaptive ensemble learning model , author =. Scientific Reports , volume =. 2022 , publisher =. doi:10.1038/s41598-022-26023-1 , url =

    work page doi:10.1038/s41598-022-26023-1 2022

  15. [15]

    Hate Speech Classifiers are Culturally Insensitive

    Lee, Nayeon and Jung, Chani and Oh, Alice. Hate Speech Classifiers are Culturally Insensitive. Proceedings of the First Workshop on Cross-Cultural Considerations in NLP (C3NLP). 2023. doi:10.18653/v1/2023.c3nlp-1.5

    work page doi:10.18653/v1/2023.c3nlp-1.5 2023

  16. [16]

    K-MHaS: A Multi-label Hate Speech Detection Dataset in Korean Online News Comment

    Lee, Jean and Lim, Taejun and Lee, Heejun and Jo, Bogeun and Kim, Yangsok and Yoon, Heegeun and Han, Soyeon Caren. K-MHaS: A Multi-label Hate Speech Detection Dataset in Korean Online News Comment. Proceedings of the 29th International Conference on Computational Linguistics. 2022. doi:10.48550/arXiv.2208.10684

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2208.10684 2022

  17. [17]

    KOLD : K orean Offensive Language Dataset

    Jeong, Younghoon and Oh, Juhyun and Lee, Jongwon and Ahn, Jaimeen and Moon, Jihyung and Park, Sungjoon and Oh, Alice. KOLD : K orean Offensive Language Dataset. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022. doi:10.18653/v1/2022.emnlp-main.744

    work page doi:10.18653/v1/2022.emnlp-main.744 2022

  18. [18]

    BEEP ! K orean Corpus of Online News Comments for Toxic Speech Detection

    Moon, Jihyung and Cho, Won Ik and Lee, Junbum. BEEP ! K orean Corpus of Online News Comments for Toxic Speech Detection. Proceedings of the Eighth International Workshop on Natural Language Processing for Social Media. 2020. doi:10.18653/v1/2020.socialnlp-1.4

    work page doi:10.18653/v1/2020.socialnlp-1.4 2020

  19. [19]

    IEEE Transactions on Artificial Intelligence (TAI) , year =

    A Review of Text Style Transfer using Deep Learning , author =. IEEE Transactions on Artificial Intelligence (TAI) , year =

  20. [20]

    Computational Linguistics , volume =

    Deep Learning for Text Style Transfer: A Survey , author =. Computational Linguistics , volume =. 2022 , address =. doi:10.1162/coli\_a\_00426 , url =

    doi:10.1162/coli 2022

  21. [21]

    arXiv preprint , volume =

    HateRephrase: Zero- and Few-Shot Reduction of Hate Intensity in Online Posts using Large Language Models , author =. arXiv preprint , volume =. 2023 , url =

    2023

  22. [22]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =

    From the Detection of Toxic Spans in Online Discussions to the Analysis of Toxic-to-Civil Transfer , author =. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages =. 2022 , publisher =. doi:10.18653/v1/2022.acl-long.259 , url =

    work page doi:10.18653/v1/2022.acl-long.259 2022

  23. [23]

    2021 , howpublished =

    KoT5: Wisenut Research Korean Text-To-Text Transfer Transformer , author =. 2021 , howpublished =

    2021

  24. [24]

    Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020) , pages =

    Temporally‑Informed Analysis of Named Entity Recognition , author =. Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics (ACL 2020) , pages =. 2020 , address =. doi:10.18653/v1/2020.acl-main.680 , url =

    work page doi:10.18653/v1/2020.acl-main.680 2020

  25. [25]

    Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP , pages =

    Why only Micro-F1? Class Weighting of Measures for Relation Classification , author =. Proceedings of NLP Power! The First Workshop on Efficient Benchmarking in NLP , pages =. 2022 , address =. doi:10.18653/v1/2022.nlppower-1.4 , url =

    work page doi:10.18653/v1/2022.nlppower-1.4 2022

  26. [26]

    Gabriel, Saadia and Puri, Isha and Xu, Xuhai and Malgaroli, Matteo and Ghassemi, Marzyeh , editor =. Can. Findings of the Association for Computational Linguistics: EMNLP 2024 , pages =. 2024 , address =. doi:10.18653/v1/2024.findings-emnlp.120 , url =

    doi:10.18653/v1/2024.findings-emnlp.120 2024

  27. [27]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages =

    Evaluating the Effectiveness of Large Language Models in Establishing Conversational Grounding , author =. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages =. 2024 , address =. doi:10.18653/v1/2024.emnlp-main.545 , url =

    doi:10.18653/v1/2024.emnlp-main.545 2024

  28. [28]

    Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , pages=

    HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection , author=. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics (ACL) , pages=. 2021 , publisher=

    2021

  29. [29]

    2025 , address=

    Bai, Zewen and Yang, Liang and Yin, Shengdi and Lu, Junyu and Zeng, Jingjie and Zhu, Haohao and Sun, Yuanyuan and Lin, Hongfei , booktitle=. 2025 , address=. doi:10.18653/v1/2025.findings-acl.532 , url=

    doi:10.18653/v1/2025.findings-acl.532 2025

  30. [30]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    Pith/arXiv arXiv

  31. [31]

    arXiv preprint arXiv:2202.11176 , year=

    A New Generation of Perspective API: Efficient Multilingual Character-level Transformers , author=. arXiv preprint arXiv:2202.11176 , year=

    Pith/arXiv arXiv

  32. [32]

    arXiv preprint arXiv:2402.15951 , year=

    Detoxllm: A framework for detoxification with explanations , author=. arXiv preprint arXiv:2402.15951 , year=

    Pith/arXiv arXiv

  33. [33]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Paradetox: Detoxification with parallel data , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  34. [34]

    arXiv preprint arXiv:2402.01888 , year=

    DetoxLLM: Large Language Models Can Detoxify Themselves , author=. arXiv preprint arXiv:2402.01888 , year=

    Pith/arXiv arXiv