Pith. sign in

REVIEW 3 major objections 6 minor 51 references

Towards Efficient and Explainable Hate Speech Detection via Model Distillation

T0 review · 3 major / 6 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read A distilled 8B language model can match a 70B teacher's explanations for hate speech while beating it on classification, the paper argues.

desk verdict A real but confounded distillation result: the 8B student beats the 70B teacher, but only because the student is fine-tuned and the teacher is few-shot, so the paper's central claim needs a label-only control. read the letter →

arxiv 2412.13698 v1 pith:TVABCXO7 submitted 2024-12-18 cs.CL

Paloma Piot , Javier Parapar This is my paper · ORCID
classification cs.CL
keywords hatespeechdetectionexplainableAIknowledgedistillationchain-of-thoughtmulti-tasklearninglargelanguagemodelscontentmoderationrationales
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 proposes distilling a 70B-parameter language model into an 8B-parameter model so that the small model both detects hate speech and explains its decisions in natural language. Using Few-Shot Chain-of-Thought prompting, the large teacher generates labels and rationales for 2,296 labelled posts, which then fine-tune the student in a multi-task framework. The central claim is that the distilled small model matches the teacher's explanation quality (human-rated correctness within one point) while outperforming it on classification (F1 0.85 versus 0.78). This matters because explainable hate speech detection becomes affordable enough for real content-moderation pipelines.

What carries the argument

The central mechanism is multi-task knowledge distillation guided by Chain-of-Thought rationales. The teacher (Llama-3-70B-Instruct) is prompted in a Few-Shot CoT setting with 12 human-authored examples to produce both a binary hate label and a natural-language rationale for each post; only the 2,296 posts where the teacher's label matches the gold label are kept. These label–rationale pairs fine-tune the student (Llama-3-8B-Instruct, 4-bit quantized via QLoRA) under the combined loss $L = \alpha L_{label} + \beta L_{rationale}$, where $L_{rationale}$ averages the per-token cross-entropy over the generated rationale sequences. The rationales are load-bearing: they carry the teacher's reasoning into the small model, and the paper argues that learning to generate them also sharpens label prediction.

What would settle it

Train the same 8B student on the same 2,296 posts using only the binary labels, with no rationales, and evaluate on MiniMetaHate Eval; if that label-only model matches the distilled model's F1 of about 0.85 and explanation quality, the claim that CoT rationale distillation drives the gain is refuted, and ordinary supervised fine-tuning would explain the result.

Watch

Extended reading notes

Core claim

The paper claims that distilling a large language model's Chain-of-Thought rationales into a smaller model yields a dual-capability system: the 8B student (Llama-3-8B-Distil-MetaHate) matches the 70B teacher on human-judged explanation completeness (100% complete for both) and correctness (91% versus 92% of posts fully correct) while surpassing it in hate speech classification (F1 0.8499 versus 0.7796). The authors attribute this to multi-task learning where the student must generate both labels and rationales, and they conclude that the distilled model 'matches the teacher model in explainability and outperforms it in classifying hate speech.'

Load-bearing premise

The central comparison assumes that pitting a fine-tuned small model against a few-shot prompted large model is a fair way to measure what distillation contributes, and the paper's own Section 5.1 admits the improvement is due to the difference in training, with no label-only fine-tuning control.

Editorial extensions

If this is right

  • If the central claim holds, a 4-bit 8B model can substitute for a 70B model in explainable hate speech detection production, cutting GPU memory from 42.5 GB to 8.1 GB and cloud inference cost by roughly a factor of seven on the reported pricing.
  • Platforms could pair a high-accuracy classifier such as MetaHateBERT with the distilled explainer, giving users natural-language justifications for every content flag while keeping the explanation model small enough to run on commodity GPUs.
  • The same Few-Shot CoT distillation recipe could generalize to other abuse detection tasks, such as toxicity, harassment, and cyberbullying, where gold labels exist and human-readable rationales are desired.
  • Because the distilled model's explanations scored 100% complete and 91% correct under human evaluation, the method offers a route to satisfying transparency obligations like the EU Digital Services Act's requirement to explain content-removal decisions.

Reading between the lines

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

  • In our reading, the headline comparison is between a fine-tuned student and a few-shot prompted teacher, so the 9% F1 gap likely reflects supervised fine-tuning on 2,296 gold labels rather than distillation per se; a label-only fine-tuning control is needed to isolate the rationales' contribution.
  • The paper's own error analysis hints the student can override the teacher's reasoning: in Table 5, the distilled model correctly flags 'married bitches' while the teacher misses it, suggesting that fine-tuning on gold labels lets the student learn patterns the few-shot teacher fails on.
  • The explanation-quality metrics rest on only 100 posts per model with three annotators, so the near-tie between teacher and distil (91% vs 92% correct) is compatible with a modest real difference; a larger human evaluation would sharpen the comparison.
  • A natural test of robustness is whether the distilled model's explanations remain faithful when paired with a different classifier, as the paper's proposed pipeline uses MetaHateBERT for labels and the distilled model for rationales, which assumes the rationale model's explanations align with another model's decisions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes a knowledge-distillation pipeline for explainable hate speech detection. A Llama-3-70B-Instruct teacher generates labels and chain-of-thought rationales for 2,993 posts using few-shot CoT prompting; the teacher's outputs are filtered to 2,296 posts where the teacher label matches the gold label, and a Llama-3-8B-Instruct student is fine-tuned on these posts with a multi-task loss combining label prediction and rationale generation. The resulting Llama-3-8B-Distil-MetaHate is evaluated on a disjoint 2,001-post subsample: it achieves F1 of 0.8499 versus 0.7796 for the teacher and 0.7467 for the base 8B model, and a human evaluation on 100 posts reports 91% correct explanations versus 92% for the teacher. The paper claims that distilled models deliver explanations of the same quality as larger models while surpassing them in classification performance.

Significance. If the central claim were fully supported, the contribution would be practically valuable: a small model that both classifies hate speech and produces human-quality rationales at substantially lower computational cost, directly relevant to DSA-style content-moderation obligations. The paper ships code, models, and prompts, and it includes a human evaluation of explanation quality, which is a strength. However, as presented, the key comparison is confounded by training regime, and the explanation-quality difference on 100 posts is within sampling noise; the significance is therefore contingent on additional experiments that separate the effect of rationale distillation from ordinary fine-tuning.

major comments (3)
  1. [§5.1, Table 2] The central claim that the distilled model 'surpasses' the teacher in classification is confounded: the teacher is evaluated in a Few-Shot CoT setting with 12 examples, while the student is fine-tuned on 2,296 gold-labelled posts with both labels and rationales. The authors themselves write that 'The improvement of the distilled model over the teacher model is due to the difference in training.' This concedes that the F1 gain (0.7796 to 0.8499) may be due to ordinary supervised fine-tuning on gold labels rather than to rationale distillation. To support the method-level claim, the paper needs a control that fine-tunes the same student on the same 2,296 labels without the rationale objective (or with the rationale objective on unrelated rationales) and compares it to the distilled model. Without this control, the empirical result stands but does not demonstrate that distillation causes the gain.
  2. [§4.1] The training set is filtered to the 2,296 of 2,993 posts where the teacher's label matched the gold label. This removes the teacher's errors from the training distribution, which could inflate the student's test performance independently of the rationales: the student is never trained on the hard cases where the teacher is wrong, while the teacher is evaluated on the unfiltered eval set. The paper should either analyse the filtering impact (e.g., train on all 2,993 teacher-generated labels, or evaluate on an unfiltered holdout with teacher labels) or justify that the filter is not responsible for the observed improvement.
  3. [§4.5, Table 4] The explanation-quality claim that the distilled model 'matches' the teacher rests on 91% vs 92% correctness on a 100-post subsample, with no confidence intervals, significance test, or analysis of annotator disagreement beyond majority voting. The difference is within sampling noise; the claim should be softened or supported with uncertainty estimates (e.g., bootstrap) and possibly a larger evaluation sample.
minor comments (6)
  1. [§3.1] The loss weights α and β in L = αL_label + βL_rationale are never specified in the experimental settings; please report these values for reproducibility.
  2. [§4.3] There is a typo: 'Llama-3-70-Instruct' should be 'Llama-3-70B-Instruct'.
  3. [§4.4] The text mentions fine-tuning MetaHateBERT on 'the same instances' used to fine-tune the LLM and reports an F1 of about 0.78, but Table 2 does not include this variant; either add it to the table or remove the claim.
  4. [§5.2, Table 3] The IAA is reported as the percentage of exact agreement for completeness and correctness, but the number of fragments per post varies; consider reporting a chance-corrected agreement measure such as Fleiss' kappa.
  5. [§5.1] The text says 'fewer than 20 examples' for the teacher's few-shot setting, but §4.3 states 12 examples; keep the numbers consistent.
  6. [§4.5] The annotation process states that annotators were presented with 'labels and explanations'; clarify whether the labels were model-predicted or gold labels, as this could affect the completeness/correctness judgments.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central results are measured against external gold labels and human judgments, not forced by the training objective or by self-citation.

full rationale

The claimed derivation chain is: (1) use Llama-3-70B with Few-Shot CoT and 12 human-authored examples to generate rationales and labels; (2) keep only the 2,296 instances where the teacher's label equals the gold label; (3) fine-tune Llama-3-8B with the multi-task loss L = αL_label + βL_rationale, where L_rationale is token-level cross-entropy against the teacher's rationales; and (4) evaluate on a held-out 2,001-post sample with gold labels, plus human assessment of 100 posts. No step makes the headline claim true by definition. The student is trained toward the teacher's rationales, so some similarity in explanation behavior is the intended objective, but the reported outcomes — F1 0.8499 vs 0.7796, and human-rated explanation correctness 91% vs 92% — are external measurements, not identities. The use of the authors' MetaHate dataset and MetaHateBERT baseline involves self-citation, but these are data and benchmark choices, not load-bearing arguments: the test labels are gold labels from 36 curated datasets, and the stronger empirical results could have failed to materialize. The main scientific weakness is a confounded comparison: the teacher is evaluated in a Few-Shot CoT setting while the student is fine-tuned on 2,296 gold-labelled posts, and the paper itself concedes that the improvement over the teacher 'is due to the difference in training' (§5.1). Without a label-only fine-tuning control, the specific attribution of the gain to rationale distillation is not established. That is a validity and experimental-design concern, not circularity: the numbers are not forced by construction. No circular step is identified.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central experimental claims rest on assumptions about label reliability, teacher rationale quality, human evaluation validity, and dataset representativeness. The paper also leaves the multi-task loss weights unreported, which is a gap because the results may depend on them.

free parameters (3)
  • alpha and beta loss weights = not reported
    The multi-task objective L = alpha*L_label + beta*L_rationale is defined in Section 3.1, but the values of alpha and beta are never stated, and no sensitivity analysis is provided.
  • Number of few-shot CoT examples = 12
    The teacher prompt includes 12 human-authored examples (Section 4.3); the paper does not report variation over this number or its effect on the generated rationales.
  • Training subset size after label filtering = 2296
    Only the 2,296 of 2,993 samples where teacher label equals gold label are used for fine-tuning (Section 4.1); this filtering rule is a data-selection choice that shapes the trained model.
assumptions (4)
  • domain assumption The adopted hate speech definition, binary hate versus non-hate and excluding offensive speech, is a valid ground truth for the MetaHate labels.
    Invoked in Section 1 and used to frame both teacher prompting and human annotation.
  • domain assumption Rationales generated by Llama-3-70B-Instruct under Few-Shot CoT prompting are sufficiently correct to serve as training targets.
    The student is trained to reproduce these rationales (Section 3.1); no automatic verification is performed before training beyond label matching.
  • domain assumption Human annotation of completeness and correctness is a reliable measure of explanation quality.
    The explanation-quality claim rests entirely on the judgments of three annotators on 100 posts (Sections 4.5 and 5.2).
  • domain assumption The two balanced subsamples of MetaHate are representative enough for the reported comparisons.
    Subsamples of about 2,000 to 3,000 posts are used for training and evaluation instead of the full 1.2M collection (Section 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Efficient and Explainable Hate Speech Detection via Model Distillation." pith.science (2026). https://pith.science/paper/TVABCXO7

@misc{pith2026241213698,
  author       = {Pith},
  title        = {Pith review of: Towards Efficient and Explainable Hate Speech Detection via Model Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TVABCXO7}},
  note         = {Machine review of arXiv:2412.13698}
}
read the original abstract

Automatic detection of hate and abusive language is essential to combat its online spread. Moreover, recognising and explaining hate speech serves to educate people about its negative effects. However, most current detection models operate as black boxes, lacking interpretability and explainability. In this context, Large Language Models (LLMs) have proven effective for hate speech detection and to promote interpretability. Nevertheless, they are computationally costly to run. In this work, we propose distilling big language models by using Chain-of-Thought to extract explanations that support the hate speech classification task. Having small language models for these tasks will contribute to their use in operational settings. In this paper, we demonstrate that distilled models deliver explanations of the same quality as larger models while surpassing them in classification performance. This dual capability, classifying and explaining, advances hate speech detection making it more affordable, understandable and actionable.

Figures

Figures reproduced from arXiv: 2412.13698 by the authors.

Figure 1
Figure 1. X’s Community Notes feature for adding post con￾text. Many ML models have limitations because they operate as black boxes, making them hard to in￾terpret. Also, these models only classify without explaining the reasons. Explaining why content is blocked or flagged as hate speech is essential for transparency and trust in online platforms [18,28]. The Digital Services Act (DSA) [9], which came into force for all plat… view at source ↗
Figure 2
Figure 2. LLM Knowledge distillation with CoT overview. 3.1 Knowledge Distillation Knowledge Distillation is a machine learning technique designed to transfer knowledge from a large pre-trained model, called teacher, to a smaller student model. The goal of knowledge distillation is to train a smaller, more efficient model to replicate the behaviour of a larger, more complex model. Next, we explain how we performed this proces… view at source ↗
Figure 3
Figure 3. Overview of our proposed approach for explaining and detecting hate speech: First, we extract rationales from an LLM using Few-Shot CoT. We then use these rationales, along with the labels, to train a small model within a multi-task learning framework. This enables the small model to explain and detect hate speech effectively [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 18 canonical work pages

  1. [1]

    AI@Meta: Llama 3 model card (2024),https://github.com/meta-llama/llama3/ blob/main/MODEL_CARD.md

    2024

  2. [2]

    Anti-Defamation League: Online hate and harassment: The ameri- can experience 2024 (2024), https://www.adl.org/resources/report/ online-hate-and-harassment-american-experience-2024

    work page 2024

  3. [3]

    In: Chung, Y.l., R\"ottger, P., Nozza, D., Talat, Z., Mostafazadeh Davani, A

    Plaza-del arco, F.M., Nozza, D., Hovy, D.: Respectful or toxic? using zero-shot learning with language models to detect hate speech. In: Chung, Y.l., R\"ottger, P., Nozza, D., Talat, Z., Mostafazadeh Davani, A. (eds.) The 7th Workshop on Online Abuse and Harms (WOAH). pp. 60–68. Association for Computational Linguistics, Toronto, Canada (Jul 2023).https:/...

    doi:10.18653/v1/2023.woah-1.6 2023

  4. [4]

    Bender, E.M., Gebru, T., McMillan-Major, A., Shmitchell, S.: On the dangers of stochastic parrots: Can language models be too big? In: Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. p. 610–623. FAccT ’21, Association for Computing Machinery, New York, NY, USA (2021). https://doi.org/10.1145/3442188.3445922

    arXiv 2021

  5. [5]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J.D., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D., Wu, J., Winter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford, A....

    2020

  6. [6]

    In: Mostafazadeh Davani, A., Kiela, D., Lambert, M., Vidgen, B., Prabhakaran, V., Waseem, Z

    Caselli, T., Basile, V., Mitrović, J., Granitzer, M.: HateBERT: Retraining BERT for abusive language detection in English. In: Mostafazadeh Davani, A., Kiela, D., Lambert, M., Vidgen, B., Prabhakaran, V., Waseem, Z. (eds.) Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021). pp. 17–25. Association for Computational Linguistics, Online (A...

    work page 2021

  7. [7]

    In: Calzolari, N., Béchet, F., Blache, P., Choukri, K., Cieri, C., Declerck, T., Goggi, S., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Moreno, A., Odijk, J., Piperidis, S

    Caselli, T., Basile, V., Mitrović, J., Kartoziya, I., Granitzer, M.: I feel offended, don’t be abusive! implicit/explicit messages in offensive and abusive language. In: Calzolari, N., Béchet, F., Blache, P., Choukri, K., Cieri, C., Declerck, T., Goggi, S., Isahara, H., Maegaard, B., Mariani, J., Mazo, H., Moreno, A., Odijk, J., Piperidis, S. (eds.) Proce...

    work page 2020

  8. [8]

    ACM Comput

    Caton, S., Haas, C.: Fairness in machine learning: A survey. ACM Comput. Surv. 56(7) (apr 2024).https://doi.org/10.1145/3616865

    doi:10.1145/3616865 2024

Show all 51 references
  1. [9]

    eu/strategy-and-policy/priorities-2019-2024/europe-fit-digital-age/ digital-services-act_en, accessed: 03/09/2024

    Commision, E.: The digital services act (2024), https://commission.europa. eu/strategy-and-policy/priorities-2019-2024/europe-fit-digital-age/ digital-services-act_en, accessed: 03/09/2024

    2024

  2. [10]

    Proceedings of the ICWSM 2023 17, 1014–1023 (Jun 2023).https://doi.org/10.1609/icwsm.v17i1.22209

    Das, M., Raj, R., Saha, P., Mathew, B., Gupta, M., Mukherjee, A.: Hatemm: A multi-modal dataset for hate video classification. Proceedings of the ICWSM 2023 17, 1014–1023 (Jun 2023).https://doi.org/10.1609/icwsm.v17i1.22209

    2023 doi

  3. [11]

    Proceedings of the ICWSM 201711(1), 512–515 (May 2017).https://doi.org/10.1609/icwsm.v11i1.14955

    Davidson,T.,Warmsley,D.,Macy,M.,Weber,I.:Automatedhatespeechdetection and the problem of offensive language. Proceedings of the ICWSM 201711(1), 512–515 (May 2017).https://doi.org/10.1609/icwsm.v11i1.14955

    2017 doi

  4. [12]

    In: Proceedings of the 39th ACM/SIGAPP Symposium on Ap- plied Computing

    De Oliveira, A.B., Baptista, C.d.S., Firmino, A.A., De Paiva, A.C.: A large lan- guage model approach to detect hate speech in political discourse using multiple language corpora. In: Proceedings of the 39th ACM/SIGAPP Symposium on Ap- plied Computing. p. 1461–1468. SAC ’24, A...

    2024

  5. [13]

    Proceedings of the ICWSM 2018 12(1) (Jun 2018).https://doi.org/10.1609/icwsm.v12i1.15041

    ElSherief, M., Kulkarni, V., Nguyen, D., Yang Wang, W., Belding, E.: Hate lingo: A target-based linguistic analysis of hate speech in social media. Proceedings of the ICWSM 2018 12(1) (Jun 2018).https://doi.org/10.1609/icwsm.v12i1.15041

    2018 doi

  6. [14]

    Proceedings of the ICWSM 201812(1) (Jun 2018).https://doi.org/10.1609/icwsm.v12i1.15038

    ElSherief, M., Nilizadeh, S., Nguyen, D., Vigna, G., Belding, E.: Peer to peer hate: Hate speech instigators and their targets. Proceedings of the ICWSM 201812(1) (Jun 2018).https://doi.org/10.1609/icwsm.v12i1.15038

    2018 doi

  7. [15]

    Proceedings of the ICWSM 201812(1) (Jun 2018)

    Founta, A., Djouvas, C., Chatzakou, D., Leontiadis, I., Blackburn, J., Stringhini, G., ..., Kourtellis, N.: Large scale crowdsourcing and characterization of twitter abusive behavior. Proceedings of the ICWSM 201812(1) (Jun 2018). https:// doi.org/10.1609/icwsm.v12i1.14991

    2018 doi

  8. [16]

    Mathematics 11(24), 5004 (Dec 2023)

    García-Díaz, J.A., Pan, R., Valencia-García, R.: Leveraging zero and few-shot learning for enhanced model generality in hate speech detection in spanish and english. Mathematics 11(24), 5004 (Dec 2023). https://doi.org/10.3390/ math11245004

    2023

  9. [17]

    In: Fišer, D., Huang, R., Prabhakaran, V., Voigt, R., Waseem, Z., Wernimont, J

    de Gibert, O., Perez, N., García-Pablos, A., Cuadros, M.: Hate speech dataset from a white supremacy forum. In: Fišer, D., Huang, R., Prabhakaran, V., Voigt, R., Waseem, Z., Wernimont, J. (eds.) Proceedings of the 2nd Workshop on Abusive Language Online (ALW2). pp. 11–20. Asso...

    2018 doi

  10. [18]

    Yale University Press (2018).https: //doi.org/10.12987/9780300235029

    Gillespie, T.: Custodians of the Internet: Platforms, Content Moderation, and the Hidden Decisions That Shape Social Media. Yale University Press (2018).https: //doi.org/10.12987/9780300235029

    2018 doi

  11. [19]

    Gupta, S., Priyadarshi, P., Gupta, M.: Hateful comment detection and hate target typepredictionforvideocomments.In:Proceedingsofthe32ndACMInternational Conference on Information and Knowledge Management. p. 3923–3927. CIKM ’23, Association for Computing Machinery, New York, NY...

    2023

  12. [20]

    In: Moosavi, N.S., Gurevych, I., Fan, A., Wolf, T., Hou, Y., Marasović, A., Ravi, S

    He, H., Shi, X., Mueller, J., Zha, S., Li, M., Karypis, G.: Distiller: A systematic study of model distillation methods in natural language processing. In: Moosavi, N.S., Gurevych, I., Fan, A., Wolf, T., Hou, Y., Marasović, A., Ravi, S. (eds.) Proceedings of the Second Worksho...

    2021 doi

  13. [21]

    In: Proceedings of the ICWSM 2023

    Hickey, D., Schmitz, M., Fessler, D., Smaldino, P.E., Muric, G., Burghardt, K.: Auditing elon musk’s impact on hate speech and bots. In: Proceedings of the ICWSM 2023. pp. 1133–1137. AAAI (2023).https://doi.org/10.1609/icwsm. v17i1.22222

    2023 doi

  14. [22]

    In: Rogers, A., Boyd-Graber, J., Okazaki, N

    Ho, N., Schmid, L., Yun, S.Y.: Large language models are reasoning teachers. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual MeetingoftheAssociationforComputationalLinguistics(Volume1:LongPapers). pp. 14852–14882. Association for Computationa...

    2023 doi

  15. [23]

    In: Rogers, A., Boyd-Graber, J., Okazaki, N

    Hsieh, C.Y., Li, C.L., Yeh, C.k., Nakhost, H., Fujii, Y., Ratner, A., Krishna, R., Lee,C.Y.,Pfister,T.:Distillingstep-by-step!outperforminglargerlanguagemodels with less training data and smaller model sizes. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Findings of the ...

  16. [24]

    Trauma, violence, & abuse 24(4), 2598–2615 (2023)

    Kansok-Dusche,J.,Ballaschk,C.,Krause,N.,Zeißig,A.,Seemann-Herz,L.,Wachs, S., Bilz, L.: A systematic review on hate speech among children and adolescents: definitions, prevalence, and overlap with related phenomena. Trauma, violence, & abuse 24(4), 2598–2615 (2023)

    2023

  17. [25]

    Kumarage, T., Bhattacharjee, A., Garland, J.: Harnessing artificial intelligence to combat online hate: Exploring the challenges and opportunities of large language models in hate speech detection (2024),https://arxiv.org/abs/2403.08035

    2024 arXiv

  18. [26]

    arXiv preprint arXiv:1910.09700 (2019)

    Lacoste, A., Luccioni, A., Schmidt, V., Dandres, T.: Quantifying the carbon emis- sions of machine learning. arXiv preprint arXiv:1910.09700 (2019)

    2019 arXiv

  19. [27]

    Lees,A.,Tran,V.Q.,Tay,Y.,Sorensen,J.,Gupta,J.,Metzler,D.,Vasserman,L.:A new generation of perspective api: Efficient multilingual character-level transform- ers.In:Proceedingsofthe28thACMSIGKDDConferenceonKnowledgeDiscovery and Data Mining. p. 3197–3207. KDD ’22, Association f...

    2022

  20. [28]

    In: Korhonen, A., Traum, D., Màrquez, L

    Liu, H., Yin, Q., Wang, W.Y.: Towards explainable NLP: A generative explanation framework for text classification. In: Korhonen, A., Traum, D., Màrquez, L. (eds.) Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. pp. 5570–5581. Associatio...

    2019 doi

  21. [29]

    In: Rogers, A., Boyd-Graber, J., Okazaki, N

    Magister, L.C., Mallinson, J., Adamek, J., Malmi, E., Severyn, A.: Teaching small language models to reason. In: Rogers, A., Boyd-Graber, J., Okazaki, N. (eds.) Proceedings of the 61st Annual Meeting of the Association for Computational Lin- guistics (Volume 2: Short Papers). ...

    2023 doi

  22. [30]

    In: Proceed- ings of the AAAI 2020 (2020), https://api.semanticscholar.org/CorpusID: 229332119 Efficient and Explainable Hate Speech Detection via Model Distillation 17

    Mathew, B., Saha, P., Yimam, S.M., Biemann, C., Goyal, P., Mukherjee, A.: Hat- explain: A benchmark dataset for explainable hate speech detection. In: Proceed- ings of the AAAI 2020 (2020), https://api.semanticscholar.org/CorpusID: 229332119 Efficient and Explainable Hate Spee...

    2020

  23. [31]

    Meta: Introducing meta llama 3: The most capable openly available llm to date (Apr 2024),https://ai.meta.com/blog/meta-llama-3/, accessed: 24/06/2024

    2024

  24. [32]

    Nations, U.: What is hate speech? (2023),https://www.un.org/en/hate-speech/ understanding-hate-speech/what-is-hate-speech , accessed: 15/11/2023

    2023

  25. [33]

    org/abs/2403.12403

    Nirmal,A.,Bhattacharjee,A.,Sheth,P.,Liu,H.:Towardsinterpretablehatespeech detection using large language model-extracted rationales (2024),https://arxiv. org/abs/2403.12403

    2024 arXiv

  26. [34]

    Computer Modeling in Engineering & Sciences140(3), 2849–2868 (2024).https: //doi.org/10.32604/cmes.2024.049631

    Pan, R., García-Díaz, J.A., Valencia-García, R.: Comparing fine-tuning, zero and few-shot strategies with large language models in hate speech detection in english. Computer Modeling in Engineering & Sciences140(3), 2849–2868 (2024).https: //doi.org/10.32604/cmes.2024.049631

    2024

  27. [35]

    Proceedings of the International AAAI Conference on Web and Social Media18(1), 2025–2039 (May 2024).https://doi.org/10.1609/ icwsm.v18i1.31445

    Piot, P., Martín-Rodilla, P., Parapar, J.: Metahate: A dataset for unifying efforts on hate speech detection. Proceedings of the International AAAI Conference on Web and Social Media18(1), 2025–2039 (May 2024).https://doi.org/10.1609/ icwsm.v18i1.31445

    2025

  28. [36]

    Proceedings of the International AAAI Conference on Web and Social Media 12(1) (Jun 2018)

    Salminen, J., Almerekhi, H., Milenković, M., Jung, S.g., An, J., Kwak, H., Jansen, B.: Anatomy of online hate: Developing a taxonomy and machine learning mod- els for identifying and classifying hate in online news media. Proceedings of the International AAAI Conference on Web...

    2018 doi

  29. [37]

    call me sexist, but

    Samory, M., Sen, I., Kohne, J., Flöck, F., Wagner, C.: “call me sexist, but...”: Revisiting sexism detection using psychological scales and adversarial samples. Proceedings of the International AAAI Conference on Web and Social Media15, 573–584 (May 2021).https://doi.org/10.16...

    2021 doi

  30. [38]

    Sanh, V., Debut, L., Chaumond, J., Wolf, T.: Distilbert, a distilled version of bert: smaller, faster, cheaper and lighter (2020),https://arxiv.org/abs/1910.01108

    2020 arXiv

  31. [39]

    In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J

    Sap,M.,Gabriel,S.,Qin,L.,Jurafsky,D.,Smith,N.A.,Choi,Y.:Socialbiasframes: Reasoning about social and power implications of language. In: Jurafsky, D., Chai, J., Schluter, N., Tetreault, J. (eds.) Proceedings of the 58th Annual Meeting of the Association for Computational Lingu...

    2020 doi

  32. [40]

    Schwartz, R., Dodge, J., Smith, N.A., Etzioni, O.: Green ai. Commun. ACM 63(12), 54–63 (nov 2020).https://doi.org/10.1145/3381831

    2020 doi

  33. [41]

    Proceedings of the ICWSM 202110(1), 687–690 (Aug 2021).https://doi.org/10.1609/icwsm.v10i1.14811

    Silva, L., Mondal, M., Correa, D., Benevenuto, F., Weber, I.: Analyzing the targets of hate in online social media. Proceedings of the ICWSM 202110(1), 687–690 (Aug 2021).https://doi.org/10.1609/icwsm.v10i1.14811

    2021 doi

  34. [42]

    In: Proceed- ings of the Fourteenth National Conference on Artificial Intelligence and Ninth Conference on Innovative Applications of Artificial Intelligence

    Spertus, E.: Smokey: automatic recognition of hostile messages. In: Proceed- ings of the Fourteenth National Conference on Artificial Intelligence and Ninth Conference on Innovative Applications of Artificial Intelligence. p. 1058–1065. AAAI’97/IAAI’97, AAAI Press (1997)

    1997

  35. [43]

    pewresearch.org/internet/2021/01/13/the-state-of-online-harassment ,ac- cessed: 03/01/2024

    Vogels, E.A.: The state of online harassment (Jan 2021), https://www. pewresearch.org/internet/2021/01/13/the-state-of-online-harassment ,ac- cessed: 03/01/2024

    2021

  36. [44]

    Wang, P., Chan, A., Ilievski, F., Chen, M., Ren, X.: Pinto: Faithful language rea- soning using prompt-generated rationales (2023),https://arxiv.org/abs/2211. 01562

    2023

  37. [45]

    Wang, Y.S., Chang, Y.: Toxicity detection with generative prompt-based inference (2022), https://arxiv.org/abs/2205.12390 18 Paloma Piot and Javier Parapar

    2022 arXiv

  38. [46]

    In: Andreas, J., Choi, E., Lazaridou, A

    Waseem, Z., Hovy, D.: Hateful symbols or hateful people? predictive features for hate speech detection on Twitter. In: Andreas, J., Choi, E., Lazaridou, A. (eds.) Proceedings of the NAACL Student Research Workshop. pp. 88–93. Association for Computational Linguistics, San Dieg...

    2016 doi

  39. [47]

    NIPS ’22, Curran Associates Inc., Red Hook, NY, USA (2024)

    Wei, J., Wang, X., Schuurmans, D., Bosma, M., Ichter, B., Xia, F., Chi, E.H., Le, Q.V., Zhou, D.: Chain-of-thought prompting elicits reasoning in large language models.In:Proceedingsofthe36thInternationalConferenceonNeuralInformation Processing Systems. NIPS ’22, Curran Associ...

    2024

  40. [48]

    In: Bouamor, H., Pino, J., Bali, K

    Yang, Y., Kim, J., Kim, Y., Ho, N., Thorne, J., Yun, S.Y.: HARE: Explainable hate speech detection with step-by-step reasoning. In: Bouamor, H., Pino, J., Bali, K. (eds.) Findings of the Association for Computational Linguistics: EMNLP 2023. pp. 5490–5505. Association for Comp...

    2023 doi

  41. [49]

    In: Burstein, J., Doran, C., Solorio, T

    Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N., Kumar, R.: Pre- dicting the type and target of offensive posts in social media. In: Burstein, J., Doran, C., Solorio, T. (eds.) Proceedings of the 2019 Conference of the North American Chapter of the Association f...

    2019 doi

  42. [50]

    Proceedings of the AAAI Conference on Artificial Intelligence38(19), 21779–21787 (Mar 2024)

    Zhang, J., Wu, Q., Xu, Y., Cao, C., Du, Z., Psounis, K.: Efficient toxic content detection by bootstrapping and distilling large language models. Proceedings of the AAAI Conference on Artificial Intelligence38(19), 21779–21787 (Mar 2024). https://doi.org/10.1609/aaai.v38i19.30178

    2024 doi

  43. [2023]

    8003–8017

    pp. 8003–8017. Association for Computational Linguistics, Toronto, Canada (Jul 2023).https://doi.org/10.18653/v1/2023.findings-acl.507

    2023 doi

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.