Leveraging the Potential of Prompt Engineering for Hate Speech Detection in Low-Resource Languages

Anomadarshi Barua (George Mason University); Management); Ruhina Tabasshum Prome (Bangladesh Institute of Governance; Tarikul Islam Tamiti (George Mason University)

REVIEW 3 major objections 6 minor 66 references

Leveraging the Potential of Prompt Engineering for Hate Speech Detection in Low-Resource Languages

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Metaphor prompting pushes Bengali hate-speech detection to 95.89% F1, beating deep-learning baselines.

desk verdict The observed effect is plausible—metaphor relabeling reduces Llama2's refusals and improves F1—but the headline 95.89 Bengali F1 is not trustworthy until the paper shows the fine-tuning and evaluation subsamples are disjoint. read the letter →

arxiv 2506.23930 v1 pith:YF5MVWDQ submitted 2025-06-30 cs.CL cs.AI

Ruhina Tabasshum Prome (Bangladesh Institute of Governance , Management) , Tarikul Islam Tamiti (George Mason University) , Anomadarshi Barua (George Mason University) This is my paper

classification cs.CLcs.AI

keywords hatespeechdetectionlow-resourcelanguagesBengalipromptengineeringjailbreakingmetaphorpromptingLlama2-7Benvironmentalimpact

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a simple prompt rewrite can make a general-purpose large language model a competitive hate-speech detector in Bengali and other languages with scarce labeled data. Its new 'metaphor prompting' replaces the trigger word 'hate' with neutral metaphor pairs such as 'summer' and 'winter,' so the model's safety guardrails do not block the answer. On a 500-comment Bengali sample, the authors report F1 rising from 73.36% without the metaphor to 95.89% with the summer-winter pair, above both their deep-learning baselines and the published BD-SHS benchmark. If the result holds, low-resource languages could get accurate content moderation without large annotated corpora, at the cost of a translation step and a small fine-tuning set.

What carries the argument

The central object is metaphor prompting, a jailbreaking prompt template that substitutes emotionally charged class labels with neutral metaphor pairs (red/green, rose-thorn, honey-venom, summer-winter) and instructs the model to answer with the metaphor so that its built-in refusal behavior is not triggered; the metaphor is then mapped back to a hate/not-hate label. The argument also relies on a chain-of-translation step in which Bengali, Hindi, and German inputs are machine-translated into English before prompting, and on LoRA fine-tuning of Llama2-7B on 500 combined-language entries.

What would settle it

Locate the exact 500 instances in the fine-tuning subsample and the 500 per-language evaluation instances. Any overlap between the two would mean the reported F1 numbers include memorized answers; a clean check would rerun the summer-winter prompt on a Bengali test set that was never shown during fine-tuning and compare its F1 with the reported 95.89%.

Watch

Extended reading notes

Core claim

The central claim is that metaphor prompting—replacing the sensitive word 'hate' with neutral metaphorical pairs such as red/green, rose-thorn, honey-venom, and summer-winter in the classification instruction—lets a LoRA-fine-tuned Llama2-7B model bypass its own safety refusals and classify hate speech accurately in low-resource languages. On a 500-comment Bengali sample, the F1 rises from 73.36% without the metaphor to 95.89% with summer-winter, which the authors say surpasses the current BD-SHS benchmark and their deep-learning baselines. With rose-thorn, English reaches 77.74% and Hindi 87.15%; summer-winter gives German 82.62%. The authors maintain that this makes metaphor prompting comparable or superior to CNN, MLP, and BiGRU classifiers while using a smaller fine-tuning set and, in most comparisons, a lower impact factor.

Load-bearing premise

The paper never states that the 500 comments used to fine-tune Llama2-7B are disjoint from the 500 comments per language used to evaluate the prompts; if the same comments appear in both sets, the reported F1 scores reflect memorization rather than detection.

Editorial extensions

If this is right

On the Bengali dataset, the summer-winter metaphor raises the reported F1 from 73.36% to 95.89%, surpassing the best deep-learning baseline (BiGRU with Word2Vec at 89.60%) and the BD-SHS benchmark.
The best metaphor pair differs by language: summer-winter gives the highest reported F1 for Bengali (95.89%) and German (82.62%), while rose-thorn gives the highest for English (77.74%) and Hindi (87.15%).
Metaphor prompting yields lower impact-factor scores than the other prompting strategies because it avoids long refusal text and extra prompt scaffolding, reducing normalized computation time, electricity, and CO2.
A single Llama2-7B fine-tuned on a 500-entry combined multilingual set and prompted with metaphors can match or beat language-specific deep-learning models, meaning no per-language model training is needed.

Reading between the lines

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

A testable extension is to check whether the 500-per-language evaluation instances are disjoint from the 500 fine-tuning instances; the paper does not state this, and any overlap would mean the reported gains partly reflect memorization rather than detection.
The metaphor words themselves are arbitrary label swaps, so the method's effect could be probed by randomizing the metaphor pairs across repeated runs; if performance is robust to which pair is used, the explanation is likely neutralization of trigger words, not any particular metaphor's semantics.
Because all non-English inputs are translated to English before prompting, the reported cross-lingual gains may partly come from translation quality; prompting a Bangla-capable model directly would separate translation effects from metaphor effects.
The comparison with deep-learning baselines is not symmetric: baselines train on full per-language datasets, while the LLM is fine-tuned on 500 combined entries and evaluated on 500 per language, so a controlled replication should match training data sizes and splits.

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.

Desk Editor's Note

read the letter

The paper reports a real, plausible effect: Llama2-7B refuses to classify text when the prompt says 'hate speech', and substituting metaphor pairs (red/green, summer/winter) gets it to comply, raising Bengali F1 from 73.36 to 95.89 on a 500-comment sample. The direction is visible in the tables and consistent with the known label-substitution jailbreak family. Calling it a 'significant departure from existing jailbreaking methods' oversells it, but the measurements themselves are new.

What the paper does well: it runs six prompting strategies across four languages, adds an environmental impact factor (a useful axis even if the weights in Eq. 3 are hand-chosen), and honestly notes that the deep learning baselines are trained per language while the LLM is fine-tuned on a combined 500-entry subsample. It also admits LLMs carry a higher environmental cost. Those touches show the authors are engaging with the problem seriously.

The soft spots are load-bearing. First, Section IV.C.2 says 500 entries are subsampled from the combined dataset for fine-tuning; Section VII.B says evaluation is restricted to 500 instances per language. The paper never states that those per-language evaluation instances come from a test split disjoint from the fine-tuning subsample. If they overlap, the reported F1 reflects memorization, not detection, and every headline comparison collapses. This is the biggest issue. Second, the 'surpasses SOTA' claim in Section VII.C compares against a published benchmark score computed on a different (likely full) test set, not on the same 500 instances. Third, the best metaphor per language is selected post hoc on the test set (Table VIII), which inflates the improvement. There are no error bars, no code, and no data artifacts, so I could not verify the raw numbers.

That said, the underlying phenomenon likely survives a clean rerun: relabeling sensitive categories reduces refusals, and a 500-instance sample with matched test sets would settle it. This is exactly the kind of paper that deserves a serious referee—the question is well-motivated, the experiments are broad, and the flaws are fixable with a stated split, matched baselines, and preregistered metaphor selection. I would send it to review with the expectation of major revision, but I would not trust the 95.89 F1 as reported until the data-handling issue is resolved.

Referee Report

3 major / 6 minor

Summary. The paper presents an empirical study of prompt-engineering strategies for binary hate speech detection with Llama2-7B across Bengali, English, German, and Hindi, introducing 'metaphor prompting' in which hate/not-hate labels are replaced by metaphor pairs such as red/green or summer/winter. It compares F1 scores and an environmental impact factor (IF) against MLP, CNN, and BiGRU baselines with GloVe, Word2Vec, and FastText embeddings. The central claim is that metaphor prompting yields comparable or superior outcomes, including a Bengali F1 of 95.89 that is said to surpass the BD-SHS benchmark.

Significance. If the evaluation were performed on clean, held-out data disjoint from fine-tuning, the result would be a practically relevant demonstration that prompt-level relabeling can bypass safety refusals and yield competitive hate speech detection in low-resource settings, with a useful environmental-impact comparison. The paper contains a broad set of experiments, four datasets, several prompting variants, and explicit environmental accounting via CodeCarbon. However, the current manuscript does not establish the central claim because the fine-tuning and evaluation subsamples are not demonstrated to be disjoint, and the per-language reported maxima are selected over four metaphor pairs. The breadth and the IF framework are strengths, but the load-bearing comparison needs to be re-run or documented before the claims can be accepted.

major comments (3)

[§IV.C.2, §VII.B, §IV.C.3] The fine-tuning step in §IV.C.2 subsamples 500 entries from the combined dataset and fine-tunes Llama2-7B on them, while §VII.B states that prompting experiments are restricted to 500 instances from each language dataset; nowhere do the authors state that these per-language evaluation instances are disjoint from the 500 fine-tuning instances. The '80-20 split' mentioned in §IV.C.3 is described only for the fine-tuning pipeline and does not establish a held-out prompt-evaluation set. Because the headline F1 of 95.89 in §VII.C (Table VIII V37) and the 'surpasses the current SOTA benchmark' statement would be invalid under memorization, the authors should either specify the exact data split with instance-level identifiers (e.g., hashes), or re-run the evaluation on a held-out test set per language that excludes all 500 fine-tuning instances, and report results separately for each split.
[§VII.C, Table VIII] The per-language headline improvements are reported as the maximum over the four metaphor pairs (Bengali 95.89 with summer-winter, English 77.74 with rose-thorn, Hindi 87.15 with rose-thorn, German 82.62 with summer-winter). Reporting the best-of-four outcome as the method's performance is a form of test-set overfitting; with only four variants and no fixed a priori selection rule or multiple-comparison adjustment, the 22.53-point Bengali gain and the 'supersedes SOTA' claim overstate the expected effect. The authors should present the full distribution across V34-V37 and select the metaphor pair by a pre-specified rule (or justify why per-language selection is legitimate) before comparing with the baselines.
[§VII.A, §VII.B, Table I vs Table VIII] The abstract's 'comparable or superior outcomes' versus CNN, MLP, and BiGRU is not supported by the current experimental protocol: Table I baselines are trained per language on the full datasets, whereas the prompted Llama2 is fine-tuned on 500 combined entries and evaluated on 500 per language, and the text does not state that both are scored on the same test instances. The authors should evaluate the baselines and the prompted LLM on the same held-out test instances (ideally the same 500-instance subsets), or at minimum report class distributions and label overlap, so the comparison in §VII.D is apples-to-apples.

minor comments (6)

[§I, §V.A] The number of prompting strategies is inconsistent: the abstract says six (zero-shot, refusal suppression, flattering, multi-shot, role, metaphor), but §V.A says 'five conventional promptings' and then also describes ICL and offering rewards, making seven conventional categories plus metaphor. Recommend unifying the taxonomy and the counts in the abstract, introduction, and related-work summary.
[§IV.C.3] The text states that temperature is set to 0 for deterministic output while also setting do_sample=True and top_k=10; with do_sample=True, the generation is not fully deterministic. Please clarify whether sampling is actually disabled or why determinism is claimed.
[Table VIII] The 'Without Metaphor' row V33 appears to be a zero-shot baseline, but its prompt structure is not defined in the text; clarify how V33 differs from V1 in Table II so the reader can interpret the 73.36 baseline and the subsequent gains.
[§VII.C] F1 scores are reported to two decimals without confidence intervals or significance tests; given the 500-instance evaluation, small differences may not be stable. Provide variance estimates or bootstrap intervals for the key comparisons.
[Throughout] There are several typos and terminology slips: 'Llma2-7B' in Section I, 'hatespeeech' in the ICL prompt example in §V.A, and inconsistent use of 'prompting strategies' versus 'promptings'. A careful proofread is needed.
[§III.A] The BD-SHS dataset is described as balanced with 48.04% hate speech, but the paper does not state the class distribution of the 500-instance per-language subsamples used for prompting; this information is needed to interpret the weighted F1 scores and to assess the comparability of baselines and prompted-LLM evaluations.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity (score 2): this is an empirical prompt-variation study with no derivation that reduces to its inputs; the lone self-citation is non-load-bearing background, and the train/eval-overlap ambiguity plus best-of-four metaphor reporting are methodological risks, not demonstrated circular reductions.

full rationale

This paper is an empirical prompt-variation study; there is no derivation chain in which a claimed prediction is equivalent to its inputs by construction. The central result — that renaming the classes with metaphor pairs (red/green, rose/thorn, honey/venom, summer/winter) in the prompt bypasses Llama2-7B's refusals and raises F1 — is a measured intervention outcome (Table VIII), not a quantity derived from fitted parameters; the F1 values record how well the fine-tuned model applies the annotation rule embedded in the prompt to 500-instance samples, which is a genuine empirical quantity. The paper's only self-citation ([42], which lists co-author T. I. Tamiti) supports a background remark about code-mixing in social media text and carries none of the paper's load, so it does not constitute circularity under the hard rules. The two substantive caveats are methodological, not circular: (i) Section IV.C.2 states that 'We subsample 500 entries from the combined dataset and provide the subsampled data to Llama2-7B to fine-tune it,' while Section VII.B says experiments are 'restricted to 500 instances at once from each language dataset,' and no passage states that the per-language evaluation instances are disjoint from that fine-tuning subsample — overlap would make the headline scores reflect memorization, but the overlap is not exhibited in the text; (ii) the advertised 22.53% Bengali gain is the maximum over four metaphor variants (Table VIII V34–V37), i.e., the headline is selected on the dependent variable. Because neither issue is a reduction-by-construction that can be exhibited from the paper's own text, the circularity score stays at 2 rather than higher.

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

The paper's claims rest on unstated premises: deterministic single-run LLM outputs stand in for the model's behavior (Section IV.C.3), machine translation preserves hate speech semantics (Section IV.C.1), keyword-parsing of outputs is lossless (Section IV.C.4), and CodeCarbon's environmental estimates are comparable across runs (Section VI.B). The free parameters are the IF weights (Eq. 3), the 500-instance subsample, and the per-language metaphor choice. No new entities are invented; the metaphor pairs are output labels, not postulated mechanisms.

free parameters (3)

Impact factor weights = 0.4 (time), 0.3 (electricity), 0.3 (CO2)
Eq. 3 assigns these weights by hand; no sensitivity analysis is reported, so IF rankings across prompts are partly an artifact of the chosen weights.
Fine-tuning/evaluation subsample size = 500
Section IV.C.2 and VII.B restrict all Llama2 experiments to 500 instances per condition for resource reasons; headline scores depend on this sample.
Per-language metaphor pair selection = summer-winter (Bengali, German), rose-thorn (English, Hindi)
Table VIII shows the best pair varies by language; the reported 'metaphor prompting' gain is the test-set maximum over four pairs, selected after seeing results.

assumptions (5)

domain assumption Single-run outputs from Llama2-7B at temperature=0 and top_k=10 are deterministic and representative of the model's classification behavior.
Section IV.C.3 sets these generation parameters; no repeated runs or variance estimates are reported, so all F1/IF comparisons assume one output per prompt is sufficient.
domain assumption Google Translate preserves the hate/not-hate semantics of Bengali, Hindi, and German comments when translated to English.
Section IV.C.1 ('Chain of translation') translates all non-English data before prompting; if translation shifts label-relevant meaning, the reported cross-language F1 is an artifact.
domain assumption The keyword-position output parser faithfully recovers the model's intended class.
Section IV.C.4 describes assigning the label of the earliest keyword found in the output; no error analysis of parsing failures is provided.
domain assumption Fine-tuning on 500 mixed-language examples transfers to each of the four languages.
Section IV.C.2 mixes all four languages into one 500-entry fine-tuning set; the paper does not report per-language validation during fine-tuning.
domain assumption CodeCarbon's CO2 and electricity estimates are accurate enough for the cross-prompt IF comparisons.
Section VI.B uses CodeCarbon; no calibration or hardware variance analysis is given.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Leveraging the Potential of Prompt Engineering for Hate Speech Detection in Low-Resource Languages." pith.science (2026). https://pith.science/paper/YF5MVWDQ

@misc{pith2026250623930,
  author       = {Pith},
  title        = {Pith review of: Leveraging the Potential of Prompt Engineering for Hate Speech Detection in Low-Resource Languages},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YF5MVWDQ}},
  note         = {Machine review of arXiv:2506.23930}
}

abstract

The rapid expansion of social media leads to a marked increase in hate speech, which threatens personal lives and results in numerous hate crimes. Detecting hate speech presents several challenges: diverse dialects, frequent code-mixing, and the prevalence of misspelled words in user-generated content on social media platforms. Recent progress in hate speech detection is typically concentrated on high-resource languages. However, low-resource languages still face significant challenges due to the lack of large-scale, high-quality datasets. This paper investigates how we can overcome this limitation via prompt engineering on large language models (LLMs) focusing on low-resource Bengali language. We investigate six prompting strategies - zero-shot prompting, refusal suppression, flattering the classifier, multi-shot prompting, role prompting, and finally our innovative metaphor prompting to detect hate speech effectively in low-resource languages. We pioneer the metaphor prompting to circumvent the built-in safety mechanisms of LLMs that marks a significant departure from existing jailbreaking methods. We investigate all six different prompting strategies on the Llama2-7B model and compare the results extensively with three pre-trained word embeddings - GloVe, Word2Vec, and FastText for three different deep learning models - multilayer perceptron (MLP), convolutional neural network (CNN), and bidirectional gated recurrent unit (BiGRU). To prove the effectiveness of our metaphor prompting in the low-resource Bengali language, we also evaluate it in another low-resource language - Hindi, and two high-resource languages - English and German. The performance of all prompting techniques is evaluated using the F1 score, and environmental impact factor (IF), which measures CO$_2$ emissions, electricity usage, and computational time.

Figures

Figures reproduced from arXiv: 2506.23930 by the authors.

**Figure 2.** The detailed methodology to create deep learning-based baseline models to detect hate speech separately from multilingual [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** The detailed methodology to evaluate conventional prompting and our proposed metaphor prompting on LLMs - this [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: Metaphor prompting jailbreaks safety guardrails of [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 6.** Figure 6: Comparison of metaphor prompting with conventional [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]

**Figure 5.** Figure 5: Comparison of metaphor prompting with conventional [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

66 extracted references · 46 canonical work pages

[1]

Available: https://www.un.org/en/hate-speech/ understanding-hate-speech/what-is-hate-speech#:~:text=To% 20provide%20a%20unified%20framework,person%20or%20a% 20group%20on

[Online]. Available: https://www.un.org/en/hate-speech/ understanding-hate-speech/what-is-hate-speech#:~:text=To% 20provide%20a%20unified%20framework,person%20or%20a% 20group%20on

work page
[2]

Hate speech review in the context of online social networks,

N. Chetty and S. Alathur, “Hate speech review in the context of online social networks,” Aggression and violent behavior , vol. 40, pp. 108–118, 2018

work page 2018
[3]

Hossain, Oct 2019

I. Hossain, Oct 2019. [Online]. Available: https://archive.dhakatribune.com/bangladesh/nation/2019/10/21/ link-between-social-media-and-communal-violence

work page 2019
[4]

Indian mob kills man over beef eating rumour,

S. Khalid, “Indian mob kills man over beef eating rumour,” Oct

work page
[5]

Business reputation and social media: A primer on threats and responses,

I. S. Horn, T. Taros, S. Dirkes, L. Hüer, M. Rose, R. Tietmeyer, and E. Constantinides, “Business reputation and social media: A primer on threats and responses,” Journal of direct, data and digital marketing practice, vol. 16, pp. 193–208, 2015

work page 2015
[6]

Exploiting cloze questions for few shot text classification and natural language inference,

T. Schick and H. Schütze, “Exploiting cloze questions for few shot text classification and natural language inference,” arXiv preprint arXiv:2001.07676, 2020

arXiv 2001
[7]

Prompt learning for low-resource multi-domain fake news detection,

Z. Peng, N. Lin, Y . Zhou, D. Zhou, and A. Yang, “Prompt learning for low-resource multi-domain fake news detection,” in 2023 international conference on asian Language Processing (IALP) . IEEE, 2023, pp. 314–319

work page 2023
[8]

Generating monolingual dataset for low resource language bodo from old books using google keep,

S. Narzary, M. Brahma, M. Narzary, G. Muchahary, P. K. Singh, A. Senapati, S. Nandi, and B. Som, “Generating monolingual dataset for low resource language bodo from old books using google keep,” in Proceedings of the Thirteenth Language Resources and Evaluation Conference, 2022, pp. 6563–6570

work page 2022

Show all 66 references

[9]

Prompt-based for low-resource tibetan text classification,

B. An, “Prompt-based for low-resource tibetan text classification,” ACM Transactions on Asian and Low-Resource Language Information Processing, vol. 22, no. 8, pp. 1–13, 2023

2023
[10]

Hate speech and of- fensive language detection in bengali,

M. Das, S. Banerjee, P. Saha, and A. Mukherjee, “Hate speech and of- fensive language detection in bengali,” arXiv preprint arXiv:2210.03479 , 2022

2022 arXiv
[11]

Tinyllm efficacy in low-resource language: An experiment on bangla text classification task,

F. N. Dehan, M. Fahim, A. Rahman, M. A. Amin, and A. A. Ali, “Tinyllm efficacy in low-resource language: An experiment on bangla text classification task,” in International Conference on Pattern Recognition . Springer, 2025, pp. 472–487

2025
[12]

Using a semi- automatic keyword dictionary for improving violent web site filtering,

R. Guermazi, M. Hammami, and A. B. Hamadou, “Using a semi- automatic keyword dictionary for improving violent web site filtering,” in 2007 Third International IEEE Conference on Signal-Image Technologies and Internet-Based System . IEEE, 2007, pp. 337–344

2007
[13]

Us and them: identifying cyber hate on twitter across multiple protected characteristics,

P. Burnap and M. L. Williams, “Us and them: identifying cyber hate on twitter across multiple protected characteristics,” EPJ Data science , vol. 5, pp. 1–15, 2016

2016
[14]

A dictionary-based approach to racism detection in dutch social media,

S. Tulkens, L. Hilte, E. Lodewyckx, B. Verhoeven, and W. Daelemans, “A dictionary-based approach to racism detection in dutch social media,” arXiv preprint arXiv:1608.08738 , 2016

2016 arXiv
[15]

A lexicon-based approach for hate speech detection,

N. D. Gitari, Z. Zuping, H. Damien, and J. Long, “A lexicon-based approach for hate speech detection,” International Journal of Multimedia and Ubiquitous Engineering , vol. 10, no. 4, pp. 215–230, 2015

2015
[16]

Hate speech detection: Challenges and solutions,

S. MacAvaney, H.-R. Yao, E. Yang, K. Russell, N. Goharian, and O. Frieder, “Hate speech detection: Challenges and solutions,” PloS one, vol. 14, no. 8, p. e0221152, 2019

2019
[17]

A web of hate: Tackling hateful speech in online social spaces,

H. M. Saleem, K. P. Dillon, S. Benesch, and D. Ruths, “A web of hate: Tackling hateful speech in online social spaces,” arXiv preprint arXiv:1709.10159, 2017

2017 arXiv
[18]

Detection of hate speech by employing support vector machine with word2vec model,

N. Sevani, I. A. Soenandi, J. Wijaya et al., “Detection of hate speech by employing support vector machine with word2vec model,” in 2021 7th International Conference on Electrical, Electronics and Information Engineering (ICEEIE). IEEE, 2021, pp. 1–5

2021
[19]

Hateful symbols or hateful people? predictive features for hate speech detection on twitter,

Z. Waseem and D. Hovy, “Hateful symbols or hateful people? predictive features for hate speech detection on twitter,” in Proceedings of the NAACL student research workshop , 2016, pp. 88–93

2016
[20]

Svm for hate speech and offensive content detection

S. Ratan, S. Sinha, and S. Singh, “Svm for hate speech and offensive content detection.” in FIRE (Working Notes), 2021, pp. 459–466

2021
[21]

A comparison of event models for naive bayes text classification,

A. McCallum, K. Nigam et al. , “A comparison of event models for naive bayes text classification,” in AAAI-98 workshop on learning for text categorization, vol. 752, no. 1. Madison, WI, 1998, pp. 41–48

1998
[22]

A survey on hate speech detection and sentiment analysis using machine learning and deep learning models,

M. Subramanian, V . E. Sathiskumar, G. Deepalakshmi, J. Cho, and G. Manikandan, “A survey on hate speech detection and sentiment analysis using machine learning and deep learning models,” Alexandria Engineering Journal, vol. 80, pp. 110–121, 2023

2023
[23]

Im- proving random forest method to detect hatespeech and offensive word,

K. Nugroho, E. Noersasongko, A. Z. Fanani, R. S. Basuki et al., “Im- proving random forest method to detect hatespeech and offensive word,” in 2019 International Conference on Information and Communications Technology (ICOIACT). IEEE, 2019, pp. 514–518

2019
[24]

But i did not mean it!—intent classification of racist posts on tumblr,

S. Agarwal and A. Sureka, “But i did not mean it!—intent classification of racist posts on tumblr,” in 2016 European Intelligence and Security Informatics Conference (EISIC) . IEEE, 2016, pp. 124–127

2016
[25]

Decision trees and random forests: Machine learning techniques to classify rare events,

S. Hegelich, “Decision trees and random forests: Machine learning techniques to classify rare events,” European policy analysis , vol. 2, no. 1, pp. 98–120, 2016

2016
[26]

Cyber hate speech on twitter: An application of machine classification and statistical modeling for policy and decision making,

P. Burnap and M. L. Williams, “Cyber hate speech on twitter: An application of machine classification and statistical modeling for policy and decision making,” Policy & internet , vol. 7, no. 2, pp. 223–242, 2015

2015
[27]

Using convolutional neural networks to classify hate-speech,

B. Gambäck and U. K. Sikdar, “Using convolutional neural networks to classify hate-speech,” in Proceedings of the first workshop on abusive language online, 2017, pp. 85–90

2017
[28]

Qutnocturnal@ hasoc’19: Cnn for hate speech and offensive content identification in hindi language,

M. A. Bashar and R. Nayak, “Qutnocturnal@ hasoc’19: Cnn for hate speech and offensive content identification in hindi language,” arXiv preprint arXiv:2008.12448, 2020

2008 arXiv
[29]

Hate speech detection using attention-based lstm,

G. L. De la Pena Sarracén, R. G. Pons, C. E. M. Cuza, and P. Rosso, “Hate speech detection using attention-based lstm,” EVALITA evaluation of NLP and speech tools for Italian , vol. 12, p. 235, 2018

2018
[30]

Detection of hate speech and offensive language in twitter data using lstm model,

A. Bisht, A. Singh, H. Bhadauria, J. Virmani, and Kriti, “Detection of hate speech and offensive language in twitter data using lstm model,” Recent trends in image and signal processing in computer vision , pp. 243–264, 2020

2020
[31]

Deep learning for hate speech detection in tweets,

P. Badjatiya, S. Gupta, M. Gupta, and V . Varma, “Deep learning for hate speech detection in tweets,” in Proceedings of the 26th international conference on World Wide Web companion , 2017, pp. 759–760

2017
[32]

A deeper look into sarcastic tweets using deep convolutional neural networks,

S. Poria, E. Cambria, D. Hazarika, and P. Vij, “A deeper look into sarcastic tweets using deep convolutional neural networks,” arXiv preprint arXiv:1610.08815, 2016

2016 arXiv
[33]

Don’t go to extremes: Revealing the excessive sensitivity and calibration limitations of llms in implicit hate speech detection,

M. Zhang, J. He, T. Ji, and C.-T. Lu, “Don’t go to extremes: Revealing the excessive sensitivity and calibration limitations of llms in implicit hate speech detection,” arXiv preprint arXiv:2402.11406 , 2024

2024 arXiv
[34]

Don’t listen to me: Understanding and exploring jailbreak prompts of large language models,

Z. Yu, X. Liu, S. Liang, Z. Cameron, C. Xiao, and N. Zhang, “Don’t listen to me: Understanding and exploring jailbreak prompts of large language models,” arXiv preprint arXiv:2403.17336 , 2024

2024 arXiv
[35]

Many-shot jailbreaking,

C. Anil, E. Durmus, M. Sharma, J. Benton, S. Kundu, J. Batson, N. Rimsky, M. Tong, J. Mu, D. Ford et al., “Many-shot jailbreaking,” Anthropic, April, 2024. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 12

2024
[36]

A survey on in-context learning,

Q. Dong, L. Li, D. Dai, C. Zheng, Z. Wu, B. Chang, X. Sun, J. Xu, and Z. Sui, “A survey on in-context learning,” arXiv preprint arXiv:2301.00234, 2022

2022 arXiv
[37]

Better zero-shot reasoning with role-play prompting,

A. Kong, S. Zhao, H. Chen, Q. Li, Y . Qin, R. Sun, and X. Zhou, “Better zero-shot reasoning with role-play prompting,” arXiv preprint arXiv:2308.07702, 2023

2023 arXiv
[38]

Respectful or toxic? using zero-shot learning with language models to detect hate speech,

F. M. Plaza-del Arco, D. Nozza, D. Hovy et al., “Respectful or toxic? using zero-shot learning with language models to detect hate speech,” in The 7th Workshop on Online Abuse and Harms (WOAH) . Association for Computational Linguistics, 2023

2023
[39]

Leveraging zero and few-shot learning for enhanced model generality in hate speech detection in spanish and english,

J. A. García-Díaz, R. Pan, and R. Valencia-García, “Leveraging zero and few-shot learning for enhanced model generality in hate speech detection in spanish and english,” Mathematics, vol. 11, no. 24, p. 5004, 2023

2023
[40]

Moderating new waves of online hate with chain-of-thought reasoning in large language models,

N. Vishwamitra, K. Guo, F. T. Romit, I. Ondracek, L. Cheng, Z. Zhao, and H. Hu, “Moderating new waves of online hate with chain-of-thought reasoning in large language models,” ArXiv, vol. abs/2312.15099, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:266551734

2023 arXiv
[41]

Hypernymy detection for low-resource languages: A study for hindi, bengali, and amharic,

A. Jana, G. Venkatesh, S. M. Yimam, and C. Biemann, “Hypernymy detection for low-resource languages: A study for hindi, bengali, and amharic,” Transactions on Asian and Low-Resource Language Information Processing, vol. 21, no. 4, pp. 1–21, 2022

2022
[42]

Milestones in bengali sentiment analysis leveraging transformer-models: Fundamentals, challenges and future directions,

S. Sengupta, S. Ghosh, P. Mitra, and T. I. Tamiti, “Milestones in bengali sentiment analysis leveraging transformer-models: Fundamentals, challenges and future directions,” arXiv preprint arXiv:2401.07847, 2024

2024 arXiv
[43]

A dataset of Hindi-English code-mixed social media text for hate speech detection,

A. Bohra, D. Vijay, V . Singh, S. S. Akhtar, and M. Shrivastava, “A dataset of Hindi-English code-mixed social media text for hate speech detection,” in Proceedings of the Second Workshop on Computational Modeling of People’s Opinions, Personality, and Emotions in Social Media...

2018
[44]

Navigating linguistic diversity: In-context learning and prompt engineering for subjectivity analysis in low-resource languages,

S. Dwivedi, S. Ghosh, and S. Dwivedi, “Navigating linguistic diversity: In-context learning and prompt engineering for subjectivity analysis in low-resource languages,” SN Computer Science , vol. 5, no. 4, p. 418, 2024

2024
[45]

Energy and policy considerations for deep learning in nlp,

E. Strubell, A. Ganesh, and A. McCallum, “Energy and policy considerations for deep learning in nlp,” ArXiv, vol. abs/1906.02243, 2019. [Online]. Available: https://api.semanticscholar.org/CorpusID:174802812

1906 arXiv
[46]

Towards climate awareness in NLP research,

D. Hershcovich, N. Webersinke, M. Kraus, J. Bingler, and M. Leippold, “Towards climate awareness in NLP research,” in Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , Y . Goldberg, Z. Kozareva, and Y . Zhang, Eds. Abu Dhabi, United Arab ...

2022
[47]

An energy-based comparative analysis of common approaches to text classification in the legal domain,

S. Gultekin, A. Globo, A. Zugarini, M. Ernandes, and L. Rigutini, “An energy-based comparative analysis of common approaches to text classification in the legal domain,” arXiv preprint arXiv:2311.01256 , 2023

2023 arXiv
[48]

Bd-shs: A benchmark dataset for learning to detect online bangla hate speech in different social contexts,

N. Romim, M. Ahmed, M. S. Islam, A. S. Sharma, H. Talukder, and M. R. Amin, “Bd-shs: A benchmark dataset for learning to detect online bangla hate speech in different social contexts,” arXiv preprint arXiv:2206.00372, 2022

2022 arXiv
[49]

A curated dataset for hate speech detection on social media text,

D. Mody, Y . Huang, and T. E. A. de Oliveira, “A curated dataset for hate speech detection on social media text,” Data in Brief , vol. 46, p. 108832, 2023

2023
[50]

Improving adversarial data collection by supporting annotators: Lessons from gahd, a german hate speech dataset,

J. Goldzycher, P. Röttger, and G. Schneider, “Improving adversarial data collection by supporting annotators: Lessons from gahd, a german hate speech dataset,” arXiv preprint arXiv:2403.19559 , 2024

2024 arXiv
[51]

Hatecheckhin: Evaluat- ing hindi hate speech detection models,

M. Das, P. Saha, B. Mathew, and A. Mukherjee, “Hatecheckhin: Evaluat- ing hindi hate speech detection models,” arXiv preprint arXiv:2205.00328, 2022

2022 arXiv
[52]

Bnlp: Natural language processing toolkit for bengali language,

S. Sarker, “Bnlp: Natural language processing toolkit for bengali language,” arXiv preprint arXiv:2102.00405 , 2021

2021 arXiv
[53]

Computers’ interpre- tations of knowledge representation using pre-conceptual schemas: an approach based on the bert and llama 2-chat models,

J. Insuasti, F. Roa, and C. M. Zapata-Jaramillo, “Computers’ interpre- tations of knowledge representation using pre-conceptual schemas: an approach based on the bert and llama 2-chat models,” Big Data and Cognitive Computing, vol. 7, no. 4, p. 182, 2023

2023
[54]

Google translate

Google, “Google translate.” [Online]. Available: https://g.co/kgs/ AzDKJYk
[55]

Chain-of-translation prompting (cotr): A novel prompting technique for low resource languages,

T. Deshpande, N. Kowtal, and R. Joshi, “Chain-of-translation prompting (cotr): A novel prompting technique for low resource languages,” arXiv preprint arXiv:2409.04512, 2024

2024 arXiv
[56]

Llama 2: Open foundation and fine-tuned chat models,

H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale et al., “Llama 2: Open foundation and fine-tuned chat models,” arXiv preprint arXiv:2307.09288, 2023

2023 arXiv
[57]

Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing,

T. Kudo and J. Richardson, “Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing,” in Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: System Demonstrations , 2018, pp. 66–71

2018
[58]

Available: https://huggingface.co/meta-llama/ Llama-2-7b-chat-hf

[Online]. Available: https://huggingface.co/meta-llama/ Llama-2-7b-chat-hf
[59]

Large language models are zero-shot reasoners,

T. Kojima, S. S. Gu, M. Reid, Y . Matsuo, and Y . Iwasawa, “Large language models are zero-shot reasoners,” Advances in neural information processing systems, vol. 35, pp. 22 199–22 213, 2022

2022
[60]

Don’t say no: Jailbreaking llm by suppressing refusal,

Y . Zhou, Z. Huang, F. Lu, Z. Qin, and W. Wang, “Don’t say no: Jailbreaking llm by suppressing refusal,”arXiv preprint arXiv:2404.16369, 2024

2024 arXiv
[61]

Flattering to deceive: The impact of sycophantic behavior on user trust in large language model,

M. V . Carro, “Flattering to deceive: The impact of sycophantic behavior on user trust in large language model,” arXiv preprint arXiv:2412.02802, 2024

2024 arXiv
[62]

Language models are few-shot learners,

T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al., “Language models are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

1901
[63]

Learning from others’ mistakes: Avoiding dataset biases without modeling them,

V . Sanh, T. Wolf, Y . Belinkov, and A. M. Rush, “Learning from others’ mistakes: Avoiding dataset biases without modeling them,” arXiv preprint arXiv:2012.01300, 2020

2012 arXiv
[64]

The butterfly effect of altering prompts: How small changes and jailbreaks affect large language model perfor- mance,

A. Salinas and F. Morstatter, “The butterfly effect of altering prompts: How small changes and jailbreaks affect large language model perfor- mance,” arXiv preprint arXiv:2401.03729 , 2024

2024 arXiv
[65]

mlco2/codecarbon: v2.4.1,

B. Courty, V . Schmidt, S. Luccioni, Goyal-Kamal, MarionCoutarel, B. Feld, J. Lecourt, LiamConnell, A. Saboni, Inimaz, supatomic, M. Léval, L. Blanche, A. Cruveiller, ouminasara, F. Zhao, A. Joshi, A. Bogroff, H. de Lavoreille, N. Laskaris, E. Abati, D. Blank, Z. Wang, A. Cato...

2024 doi
[2015]

Available: https://www.aljazeera.com/news/2015/10/1/ indian-mob-kills-man-over-beef-eating-rumour

[Online]. Available: https://www.aljazeera.com/news/2015/10/1/ indian-mob-kills-man-over-beef-eating-rumour

2015

Pith tools

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

[1] [1]

Available: https://www.un.org/en/hate-speech/ understanding-hate-speech/what-is-hate-speech#:~:text=To% 20provide%20a%20unified%20framework,person%20or%20a% 20group%20on

[Online]. Available: https://www.un.org/en/hate-speech/ understanding-hate-speech/what-is-hate-speech#:~:text=To% 20provide%20a%20unified%20framework,person%20or%20a% 20group%20on

work page

[2] [2]

Hate speech review in the context of online social networks,

N. Chetty and S. Alathur, “Hate speech review in the context of online social networks,” Aggression and violent behavior , vol. 40, pp. 108–118, 2018

work page 2018

[3] [3]

Hossain, Oct 2019

I. Hossain, Oct 2019. [Online]. Available: https://archive.dhakatribune.com/bangladesh/nation/2019/10/21/ link-between-social-media-and-communal-violence

work page 2019

[4] [4]

Indian mob kills man over beef eating rumour,

S. Khalid, “Indian mob kills man over beef eating rumour,” Oct

work page

[5] [5]

Business reputation and social media: A primer on threats and responses,

I. S. Horn, T. Taros, S. Dirkes, L. Hüer, M. Rose, R. Tietmeyer, and E. Constantinides, “Business reputation and social media: A primer on threats and responses,” Journal of direct, data and digital marketing practice, vol. 16, pp. 193–208, 2015

work page 2015

[6] [6]

Exploiting cloze questions for few shot text classification and natural language inference,

T. Schick and H. Schütze, “Exploiting cloze questions for few shot text classification and natural language inference,” arXiv preprint arXiv:2001.07676, 2020

arXiv 2001

[7] [7]

Prompt learning for low-resource multi-domain fake news detection,

Z. Peng, N. Lin, Y . Zhou, D. Zhou, and A. Yang, “Prompt learning for low-resource multi-domain fake news detection,” in 2023 international conference on asian Language Processing (IALP) . IEEE, 2023, pp. 314–319

work page 2023

[8] [8]

Generating monolingual dataset for low resource language bodo from old books using google keep,

S. Narzary, M. Brahma, M. Narzary, G. Muchahary, P. K. Singh, A. Senapati, S. Nandi, and B. Som, “Generating monolingual dataset for low resource language bodo from old books using google keep,” in Proceedings of the Thirteenth Language Resources and Evaluation Conference, 2022, pp. 6563–6570

work page 2022