Cracking the Code: Enhancing Implicit Hate Speech Detection through Coding Classification

Liangzhi Li; Lu wei; Noa Garcia; Tong Xiang; Xiao Liu

REVIEW 3 major objections 4 minor 17 references

Cracking the Code: Enhancing Implicit Hate Speech Detection through Coding Classification

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Adding a six-codetype taxonomy of rhetorical encoding tricks to prompts or to frozen-LLM embeddings improves implicit hate speech detection across Chinese and English benchmarks.

desk verdict The taxonomy construction and embedding-based gains are real, but the 'consistently improves' claim is contradicted by their own prompt-based ToxiCN results; the paper needs claim-narrowing and leakage checks. read the letter →

arxiv 2506.04693 v1 pith:G7VWJGY5 submitted 2025-06-05 cs.CL

Lu Wei , Liangzhi Li , Tong Xiang , Xiao Liu , Noa Garcia This is my paper · ORCID

classification cs.CL

keywords implicithatespeechdetectioncodetypetaxonomyrhetoricalstrategieslargelanguagemodelsembedding-basedclassificationprompt-basedcross-lingual

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

Implicit hate speech hides its target behind indirect devices: a slur compressed into an abbreviation, a group compared to an animal, a homophone swapped in to evade filters. This paper claims that large language models detect such speech more accurately when they are first shown what these encoding tricks look like. The authors define six 'codetypes' — abbreviation, metaphor, irony, pun, idiom, and argot — and supply the names, definitions, and example sentences to LLMs either in the prompt or as extra context in the embedding layer. Across one Chinese and two English datasets, the codetype-augmented setups achieve higher F1 scores than the same models without codetypes, with the largest and most consistent gains coming from the embedding-based methods. If the claim holds, content moderators get a cheap, language-agnostic way to sharpen veiled-hate detection without retraining a model.

What carries the argument

The central object is the codetype taxonomy: a set of six rhetorical encoding strategies — Abbreviation, Metaphor, Irony, Pun, Idiom, and Argot — that the paper claims cover roughly 80% of implicit hate speech in its annotated sample. The load-bearing mechanism is feeding these codetype labels, Wikipedia-derived explanations, and sample sentences to LLMs. In the prompt-based method, the codetype block is prepended to the classification instruction; in the embedding-based methods, the taxonomy is concatenated with the sentence before frozen LLMs are used as encoders, either all at once, per codetype with concatenated embeddings, or per codetype with averaged embeddings, and the multi-head attention outputs from all transformer layers are fed to a logistic regression classifier.

What would settle it

Check every test example in ToxiCN, Latent-hatred, and ISHate for exact or near-verbatim overlap with the codetype sample sentences used in the paper, such as the Abbreviation sample 'txl是社会问题，应该有人管，而不是当透明人' or 'WPWW', then rerun the prompt and embedding baselines with any overlapping samples withdrawn from the codetype information; the central claim is falsified if the F1 advantage over the no-codetype baseline disappears.

Watch

Extended reading notes

Core claim

The paper's central discovery is that implicit hate speech detection is improved by making explicit the linguistic encoding strategies that make hate implicit. Concretely, the authors propose a taxonomy of six codetypes — Abbreviation, Metaphor, Irony, Pun, Idiom, and Argot, with an Other category — built by filtering Chinese rhetorical categories through Wikipedia and annotating 200 samples from ToxiCN. They then use this taxonomy in two ways: a prompt-based method that prepends codetype names, definitions, and samples to the classification instruction, and an embedding-based method that treats LLMs as frozen encoders and supplies codetype information through three constructions — concatenating all codetypes with the sentence, concatenating per-codetype embeddings, and averaging per-codetype embeddings. On ToxiCN (Chinese), Latent-hatred (English), and ISHate (English), the embedding-augmented variants improve F1 over the no-codetype baseline, with the per-codetype matching strategies being more robust than the single-concatenation strategy. The prompt-based gains are smaller and dataset-dependent, and the paper attributes some failures to overcorrection when too much codetype information is provided.

Load-bearing premise

The evaluation assumes that the illustrative example sentences shown for each codetype are not drawn from the test portions of the evaluation datasets; if any test sentence or near-duplicate appears among those examples, the measured gains could come from memorization instead of taxonomy understanding.

Editorial extensions

If this is right

If the codetype claim is right, content moderators can improve veiled-hate detection simply by adding a six-line taxonomy to their existing prompt, without fine-tuning a model.
The embedding-based results imply that even a frozen, instruction-tuned LLM contains enough linguistic knowledge for the taxonomy to redirect its attention; no parameter updates are needed for the gain to appear.
Per-codetype matching outperforms feeding all codetypes at once, suggesting that letting the model check each encoding strategy against the sentence individually is the more reliable way to use the taxonomy.
Codetype examples act as few-shot demonstrations in prompt-based settings, but combining them with names and explanations can overcorrect the model into flagging neutral statements about sensitive groups as hateful.
The taxonomy transfers across Chinese and English, supporting the claim of cross-linguistic commonalities in how hate is encoded and pointing to a single taxonomy usable in both languages.

Reading between the lines

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

Beyond the paper, the codetype idea should transfer to other indirect harmful language, such as masked slurs in cyberbullying or dog-whistle politics; a quick test would be annotating one such corpus with the six codetypes and checking whether the same embedding boost appears.
A natural extension is an automatic codetype tagger that predicts which of the six strategies actually appears in a sentence and feeds only those labels to the model, which could reduce the overcorrection errors the paper observes with full-taxonomy inputs.
The sample-based gains suggest the taxonomy may be working partly as a retrieval scaffold — the model sees a canonical example and pattern-matches — rather than as a deep understanding of the definitions; an ablation with the sample examples removed but definitions kept would separate these two mechanisms.
If the cross-linguistic result holds, the same six codetypes could be used as an annotation schema for detecting implicit hate in third languages, testing whether the taxonomy is universal or only shared between Chinese and English.

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

Short version: the six-codetype taxonomy is a real attempt to give im-HS detection a usable linguistic handle, and the embedding-based results are genuinely better across all three datasets. But the abstract says codetypes 'consistently improve' detection, and that is not what Table 3 shows. On ToxiCN with the prompt-based method, 19 of 21 codetype configurations land below the no-codetype baseline; on Llama2-7B and Llama2-13B every single one does. The stress-test note said only one configuration beats baseline; it's actually two (Samp and Name+Samp on Baichuan2-13B, both marginal), but the point stands. Section 5.4 concedes this in a single sentence, then the conclusion quietly claims the experiments 'establish the efficacy.' That's an internal contradiction the authors need to fix, not a stylistic choice.

What's actually new: the taxonomy construction pipeline (Chinese rhetoric sources, Wikipedia filtering, word2vec similarity pruning, 200-sample annotation with Fleiss' kappa) is systematic and I haven't seen it in the cited prior work. The three embedding fusion strategies are simple but they work: the gains on Latent-hatred and ISHate, plus ToxiCN, are consistent and often large. Testing on English datasets gives the taxonomy some independent grounding, even if the Chinese origin and ToxiCN-based category selection make the circularity only partial, not absent.

Soft spots, in order. First, the claim/report mismatch above is the main issue. Second, the evaluation rests on a single 8:1:1 split with no error bars, no significance tests, and no code or data release. Some differences are large enough to be believable, but we can't rule out split luck. Third, leakage: Table 1 and Figure 1 use ToxiCN sentences as codetype examples, and the paper never states that these sentences were excluded from test splits. If they were not, the Samp-based gains could be inflated. The paper does not address this. Fourth, the taxonomy's conceptual novelty is modest—irony, metaphor, pun, and so on are ancient categories—but the contribution is the empirical packaging, not the names.

Who this is for: people working on content moderation, especially multilingual or Chinese im-HS. The overclaim makes it hard to cite as-is. Recommend: send to peer review with a request for major revision—narrow the claims to embedding-based methods and to specific configurations, add error bars or significance tests, release data/code with deduplicated splits, and state whether illustration examples overlap with test data. A serious referee can get this to a publishable state.

Referee Report

3 major / 4 minor

Summary. The paper proposes a six-category taxonomy of "codetypes" (Irony, Metaphor, Pun, Argot, Abbreviation, Idiom) claimed to characterize implicit hate speech, and it tests two ways of integrating this taxonomy into LLM-based detection: a prompt-based method that adds codetype names, explanations, and sample sentences to the instruction, and an embedding-based method that appends codetype information to the input of frozen LLMs and feeds the resulting representations to a logistic regression classifier. Experiments are run on three datasets (ToxiCN in Chinese; Latent-hatred and ISHate in English) with three models (Baichuan2-13B-Chat, Llama2-Chinese-7B-Chat, Llama2-Chinese-13B-Chat). The paper claims that codetypes consistently improve implicit hate speech detection across languages and methods.

Significance. The codetype taxonomy is a genuine descriptive contribution that organizes implicit hate speech into interpretable linguistic strategies, and the embedding-based methods show substantial and often consistent F1 gains over the no-codetype baseline, especially for the EII and EIII variants; the best result reaches 0.8091 F1 on ToxiCN. The inclusion of two English datasets provides some cross-linguistic validation, which is a strength. The paper also usefully reports negative prompt-based results in Section 5.4, although the abstract and conclusion do not reflect this nuance. The main limitation is that the headline claim of consistent improvement is contradicted by the paper's own data for the prompt-based method on ToxiCN, and the evaluation lacks statistical validation and does not document whether codetype example sentences are excluded from the test splits.

major comments (3)

[§5.4, Table 3] The central claim that codetypes "consistently improve" implicit hate speech detection is not supported by Table 3. For the prompt-based method on ToxiCN, only 1 of 21 codetype configurations (Baichuan2-13B with Samp, 0.2988 vs 0.2556 baseline) exceeds the no-codetype baseline; all seven Llama2-7B and all seven Llama2-13B prompt-based configurations fall below baseline. Section 5.4 concedes this point, but the abstract and Section 6 still assert general efficacy. The claim must be narrowed to the embedding-based methods and to the specific configurations that actually improve, and the abstract/conclusion should be revised accordingly.
[§3.1, Table 1, §5.3] The taxonomy is built from a 200-sample subset of ToxiCN, and the Samp examples used in prompts and embedding inputs are explicitly drawn from ToxiCN (e.g., Table 1's 'txl是社会问题...' and Figure 1). The paper never states whether these example sentences are excluded from the ToxiCN test split. If any Samp example overlaps with a test sentence, the gains on ToxiCN could be inflated by memorization rather than by taxonomy understanding. The authors must either confirm the exclusion of all codetype examples from test splits or re-run the ToxiCN evaluation with held-out examples before the Chinese results can be trusted.
[§5.3] All experimental results come from a single 8:1:1 split with no error bars, repeated runs, or significance tests. Many critical comparisons in Table 3 differ by less than 0.02 F1 (e.g., Llama2-7B on ISHate: 0.6625 baseline vs 0.6667 Name+Expl in EIII), so the reported improvements are not established as reliable. The authors should provide variance estimates (e.g., bootstrap confidence intervals) and significance tests for the headline comparisons, or at least discuss the magnitude of the differences relative to the lack of variance information.

minor comments (4)

[Figure 2, Figure 5, Table 5] The word 'Argot' is misspelled as 'Agort' in several places, including Figure 2 and Figure 5; this should be corrected throughout.
[§6] The section heading reads 'Conlusions' instead of 'Conclusions'.
[§3.1, Appendix A.2] The text contains corrupted glyph sequences (e.g., 'uni0000002c/uni00000055...' in Section 3.1 and Appendix A.2) that appear to be rendering errors for Chinese characters; these must be fixed to the actual characters.
[§5.3] The composition of the 'Samp' component is ambiguous: the paper should specify whether the selected samples are drawn from the training split, whether they are identical across datasets, and whether they are the same for all models and prompt variants.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: codetypes are explicitly non-equivalent to im-HS labels, the Samp examples are illustrative rather than fitted parameters, and the English benchmarks are external; the paper's internal inconsistency with Table 3 is a correctness issue, not a derivation-by-construction.

full rationale

Walking the derivation chain: the codetype taxonomy is induced from a 200-sample ToxiCN subset and from rhetorical-style literature, but Section 3.2 explicitly states the taxonomy 'is not a direct indicator of im-HS' — so a sentence's codetype is not the label it is asked to predict. The prompt/embedding inputs are the six codetype names, Wikipedia explanations, and illustrative samples (Section 5.3); these are fixed auxiliary descriptions, not per-test-instance fitted values or gold labels. No equation reduces the F1 output to a fitted constant or to the taxonomy's construction; the same codetype prompts are applied to two external English datasets (Latent-hatred, ISHate), so the improvement claim is externally checkable. The closest concern is that Table 1 and Figure 1 draw 'Samp' illustrations from ToxiCN and the taxonomy is built from ToxiCN, while the paper does not state that the 200 annotated samples or Samp strings are excluded from the ToxiCN test split; this is a potential data-leakage / robustness limitation, and the Limitations section itself concedes 'the selection of codetype samples also influences experimental performances.' Separately, the abstract and Section 1 claim codetypes 'consistently improves' detection, but Section 5.4 concedes 'the average F1 score of the prompt-based method is generally lower in ToxiCN comparing to its baseline,' and Table 3 shows only 1 of 21 prompt-based ToxiCN configurations beats baseline. These are internal-consistency/correctness failures, not circular reductions: the claim is unsupported by the paper's own numbers rather than true-by-definition. Under the circularity rubric, there is no exhibitable Eq. X = Eq. Y by construction, no fitted parameter renamed as a prediction, and no load-bearing self-citation chain; consequently the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 1 invented entities

The central claim rests on hand-picked taxonomy construction thresholds (similarity 0.9, top-six categories), a transfer assumption across languages, and the practice of preassigning all six codetypes to every sentence. The most consequential burden is the possible overlap between illustrative codetype examples and evaluation data.

free parameters (3)

codetype similarity threshold = 0.9 cosine similarity
Codetype candidates with word2vec cosine similarity above 0.9 are removed; this threshold is hand-chosen (Section 3.1).
number of final codetypes K = 6
The top six most frequent categories from 200 annotated ToxiCN samples were kept, covering about 80% of samples; K=6 is used in Methods II and III (Sections 3.1, 4.2).
training/validation/test split = 8:1:1
A fixed split with no seeds or repeats is used for all results (Section 5.3).

assumptions (4)

domain assumption Rhetorical categories in Chinese are transferable to English and to implicit hate speech.
Section 3 states the six codetypes are found in both Chinese and English datasets, but no formal criterion or external validation is provided.
domain assumption Word2vec similarity of Wikipedia explanations is a valid way to merge or filter rhetorical categories.
Section 3.1 uses cosine similarity above 0.9 to downsample 43 candidates to 10; no evaluation of this assumption is given.
domain assumption Every input sentence benefits from being paired with all six codetypes at inference time.
Section 5.3 says codetype information is preassigned for each input consisting of all six codetypes; the system does not predict which codetype is actually present.
domain assumption The benchmark labels in ToxiCN, Latent-hatred, and ISHate are treated as ground truth.
The evaluation uses the datasets' hate and no-hate labels without auditing them (Section 5.1).

invented entities (1)

Codetype taxonomy (Irony, Metaphor, Pun, Argot, Abbreviation, Idiom)
purpose: Provide six encoding strategies used to augment prompts and embeddings for implicit hate speech detection
The taxonomy is derived from a 200-sample ToxiCN subset and Wikipedia definitions; its generalization is tested only through the paper's own downstream experiments, with no external pre-registered validation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Cracking the Code: Enhancing Implicit Hate Speech Detection through Coding Classification." pith.science (2026). https://pith.science/paper/G7VWJGY5

@misc{pith2026250604693,
  author       = {Pith},
  title        = {Pith review of: Cracking the Code: Enhancing Implicit Hate Speech Detection through Coding Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G7VWJGY5}},
  note         = {Machine review of arXiv:2506.04693}
}

read the original abstract

The internet has become a hotspot for hate speech (HS), threatening societal harmony and individual well-being. While automatic detection methods perform well in identifying explicit hate speech (ex-HS), they struggle with more subtle forms, such as implicit hate speech (im-HS). We tackle this problem by introducing a new taxonomy for im-HS detection, defining six encoding strategies named codetypes. We present two methods for integrating codetypes into im-HS detection: 1) prompting large language models (LLMs) directly to classify sentences based on generated responses, and 2) using LLMs as encoders with codetypes embedded during the encoding process. Experiments show that the use of codetypes improves im-HS detection in both Chinese and English datasets, validating the effectiveness of our approach across different languages.

Figures

Figures reproduced from arXiv: 2506.04693 by the authors.

**Figure 1.** Selected examples from the ToxiCN dataset (Lu et al., 2023) that illustrate six codetypes. English translation below is provided for clarity. The keywords that encode specific types of implicit hate information are highlighted. Within the prevailing research, the majority of scholarly efforts are dedicated to categorizing macro-level concepts of HS (Jiang et al., 2023; Choi et al., 2023; Sarwar and Murdock, 2022; Al… view at source ↗

**Figure 3.** Distribution of codetype candidate categories [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

**Figure 4.** The overall framework of our prompt-based [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗

Figures from the paper (2 more)

**Figure 5.** Figure 5: The overall framework of the three proposed [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: The embedding cosine similarity between 10 [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 14 canonical work pages

[1]

In Chinese datasets, speakers often use the initial letters of Chinese pinyin for abbreviations

Abbreviation: In languages using phonetic characters, simplified forms are adopted for commonly used phrases (mostly proper nouns) and a few common words. In Chinese datasets, speakers often use the initial letters of Chinese pinyin for abbreviations. Examples: /uni0000002c/uni00000055/uni00000052/uni00000051/uni0000005c/uni00000030/uni00000048/uni0000005...

work page
[2]

Speakers liken a certain group to animals or describe a group or object using animal behav- ior, or they may use specific things or events to describe a group or object

Metaphor: It involves comparing two un- related things, creating a rhetorical transfer. Speakers liken a certain group to animals or describe a group or object using animal behav- ior, or they may use specific things or events to describe a group or object. This includes both direct and indirect forms. Examples: •宁愿小鸡守活寡？也不想小鸡沾粑粑？ Would you rather havea...

work page
[3]

Irony: The literal meaning is inconsistent with the deeper meaning, where words con- vey something different from their apparent Irony MetaphorArgot Pun AbbreviationIdiom Rhetorical*LoanwordHyperboleDeformationNone Annotator 149 26 30 16 10 6 29 4 18 1 11 Annotator 263 31 26 20 16 10 13 2 7 1 11 Annotator 378 24 22 18 17 7 15 2 5 1 11 Consensus 54 26 17 1...

work page
[4]

Examples: •艾呀，梅关系，不会疣事的，概率为淋。疣其湿没有尖锐的问题就更不担心啦，你要照顾好滋己呀，不要经常去疱吧了，找个斑上吧，疹心的。 Oh dear, don’t worry, everything will be fine, the probability is zero

Pun: Utilizing homophones or wordplay to convey an indirect meaning. Examples: •艾呀，梅关系，不会疣事的，概率为淋。疣其湿没有尖锐的问题就更不担心啦，你要照顾好滋己呀，不要经常去疱吧了，找个斑上吧，疹心的。 Oh dear, don’t worry, everything will be fine, the probability is zero. Especially with no acute issues, there’s even less to worry about. Take care of yourself and don’t spend too much time at bars. Go and f...

work page
[5]

InTENCON 2022- 2022 IEEE Region 10 Conference (TENCON), pages 1–4

Combating high variance in data-scarce im- plicit hate speech classification. InTENCON 2022- 2022 IEEE Region 10 Conference (TENCON), pages 1–4. IEEE. Kyumin Park, Myung Jae Baik, YeongJun Hwang, Yen Shin, HoJae Lee, Ruda Lee, Sang Min Lee, Je Young Hannah Sun, Ah Rah Lee, Si Yeun Yoon, Dong ho Lee, Jihyung Moon, JinYeong Bak, Kyunghyun Cho, Jong-Woo Paik...

work page 2022
[6]

Harmful Suicide Content Detection

Harmful suicide content detection.Preprint, arXiv:2407.13942. Christopher Perez and Sayar Karmakar. 2023. An nlp-assisted bayesian time-series analysis for preva- lence of twitter cyberbullying during the covid-19 pandemic.Social network analysis and mining, 13(1):51. Kyle Rapp. 2021. Social media and genocide: The case for home state responsibility.Journ...

work page Pith review arXiv 2023
[7]

In2022 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 139–146

Quantifying how hateful communities radical- ize online users. In2022 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 139–146. IEEE. Deepawali Sharma, Vedika Gupta, and Vivek Kumar Singh. 2022. Detection of homophobia & transpho- bia in malayalam and tamil: Exploring deep learn- ing methods. InInternati...

arXiv 2022
[8]

For example, The character ‘默’(silence) is composed of ‘ 黑’ (black) + ‘犬’ (dog), which is used to mock black peo- ple

Decomposition: As Chinese characters are ideograms, they can convey specific emotions through individual characters’ separation and combination. For example, The character ‘默’(silence) is composed of ‘ 黑’ (black) + ‘犬’ (dog), which is used to mock black peo- ple. Examples: •好奇纹了什么字，默吗 I wonder what character was used to cre- ate curiosity, ‘默’?

work page

Show all 17 references

[9]

Hyperbole: Intentionally magnifying and em- bellishing the characteristics of objective per- sons, events, or things to deviate from the truth, aiming to deepen the reader’s impres- sion. Examples: •我觉得，你可以在淘宝开定制戒指服务，接单给小黑做，你这等于有一个几百个小工的工厂啊 I think you could offer custom rin...
[10]

This answer may be explicit or implicit

Rhetorical Question: The speaker poses a question that seems to be directed at the reader or audience, but actually contains the author’s own answer. This answer may be explicit or implicit. This technique is mainly used to emphasize viewpoints, guide thinking, or evoke emotio...
[13]

For instance, 基佬(gay) originally appeared as a dialect in the Guangdong region

Idiom: Informal and colloquial words or phrases used by the general public, also known as dialect, vernacular, or folk idioms, including vocabulary or short phrases. For instance, 基佬(gay) originally appeared as a dialect in the Guangdong region. Examples: •东北人自强不息就 OK了，这么肥沃的黑...
[14]

Generally, loanwords do not in- clude translated words but refer to words that are absorbed with both sound and meaning

Loanword: Words borrowed from another language. Generally, loanwords do not in- clude translated words but refer to words that are absorbed with both sound and meaning. For example, 凯勒奇(Kaleqi) is the English translation of the anti-Semitic Kalergi Plan. Examples: •犹太人正在中国借助女权...
[2017]

InProceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 774–782, Taipei, Taiwan

Recognizing explicit and implicit hate speech using a weakly supervised two-path bootstrapping approach. InProceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 774–782, Taipei, Taiwan. Asian Federation of Natura...

2023 arXiv
[2021]

InProceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1438–1447, Held Online

Contextual-lexicon approach for abusive lan- guage detection. InProceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 1438–1447, Held Online. INCOMA Ltd. Pranav Venkatesh, Kami Vinton, Dhiraj Murthy, Kellen Sharp, and...

2021 arXiv
[2022]

Andrew Alexander and Hongbin Wang

Twitter users’ behavioral response to toxic replies.arXiv preprint arXiv:2210.13420. Andrew Alexander and Hongbin Wang. 2023. Topo- logical data mapping of online hate speech, mis- information, and general mental health: A large language model based study.arXiv preprint arXiv:...

2023 arXiv
[2023]

you don’t understand, this is a new war!

Rule by example: Harnessing logical rules for explainable hate speech detection.Preprint, arXiv:2307.12935. Mithun Das, Binny Mathew, Punyajoy Saha, Pawan Goyal, and Animesh Mukherjee. 2020. Hate speech in online social media.ACM SIGWEB Newsletter, 2020(Autumn):1–8. Mai ElSher...

2020 arXiv
[2024]

Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and Ece Kamar

Watching the watchers: A comparative fair- ness audit of cloud-based content moderation ser- vices.European Workshop on Algorithmic Fairness. Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and Ece Kamar. 2022. Toxigen: A large-scale machine-genera...

2022 arXiv

Pith tools

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

[1] [1]

In Chinese datasets, speakers often use the initial letters of Chinese pinyin for abbreviations

Abbreviation: In languages using phonetic characters, simplified forms are adopted for commonly used phrases (mostly proper nouns) and a few common words. In Chinese datasets, speakers often use the initial letters of Chinese pinyin for abbreviations. Examples: /uni0000002c/uni00000055/uni00000052/uni00000051/uni0000005c/uni00000030/uni00000048/uni0000005...

work page

[2] [2]

Speakers liken a certain group to animals or describe a group or object using animal behav- ior, or they may use specific things or events to describe a group or object

Metaphor: It involves comparing two un- related things, creating a rhetorical transfer. Speakers liken a certain group to animals or describe a group or object using animal behav- ior, or they may use specific things or events to describe a group or object. This includes both direct and indirect forms. Examples: •宁愿小鸡守活寡？也不想小鸡沾粑粑？ Would you rather havea...

work page

[3] [3]

Irony: The literal meaning is inconsistent with the deeper meaning, where words con- vey something different from their apparent Irony MetaphorArgot Pun AbbreviationIdiom Rhetorical*LoanwordHyperboleDeformationNone Annotator 149 26 30 16 10 6 29 4 18 1 11 Annotator 263 31 26 20 16 10 13 2 7 1 11 Annotator 378 24 22 18 17 7 15 2 5 1 11 Consensus 54 26 17 1...

work page

[4] [4]

Examples: •艾呀，梅关系，不会疣事的，概率为淋。疣其湿没有尖锐的问题就更不担心啦，你要照顾好滋己呀，不要经常去疱吧了，找个斑上吧，疹心的。 Oh dear, don’t worry, everything will be fine, the probability is zero

Pun: Utilizing homophones or wordplay to convey an indirect meaning. Examples: •艾呀，梅关系，不会疣事的，概率为淋。疣其湿没有尖锐的问题就更不担心啦，你要照顾好滋己呀，不要经常去疱吧了，找个斑上吧，疹心的。 Oh dear, don’t worry, everything will be fine, the probability is zero. Especially with no acute issues, there’s even less to worry about. Take care of yourself and don’t spend too much time at bars. Go and f...

work page

[5] [5]

InTENCON 2022- 2022 IEEE Region 10 Conference (TENCON), pages 1–4

Combating high variance in data-scarce im- plicit hate speech classification. InTENCON 2022- 2022 IEEE Region 10 Conference (TENCON), pages 1–4. IEEE. Kyumin Park, Myung Jae Baik, YeongJun Hwang, Yen Shin, HoJae Lee, Ruda Lee, Sang Min Lee, Je Young Hannah Sun, Ah Rah Lee, Si Yeun Yoon, Dong ho Lee, Jihyung Moon, JinYeong Bak, Kyunghyun Cho, Jong-Woo Paik...

work page 2022

[6] [6]

Harmful Suicide Content Detection

Harmful suicide content detection.Preprint, arXiv:2407.13942. Christopher Perez and Sayar Karmakar. 2023. An nlp-assisted bayesian time-series analysis for preva- lence of twitter cyberbullying during the covid-19 pandemic.Social network analysis and mining, 13(1):51. Kyle Rapp. 2021. Social media and genocide: The case for home state responsibility.Journ...

work page Pith review arXiv 2023

[7] [7]

In2022 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 139–146

Quantifying how hateful communities radical- ize online users. In2022 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining (ASONAM), pages 139–146. IEEE. Deepawali Sharma, Vedika Gupta, and Vivek Kumar Singh. 2022. Detection of homophobia & transpho- bia in malayalam and tamil: Exploring deep learn- ing methods. InInternati...

arXiv 2022

[8] [8]

For example, The character ‘默’(silence) is composed of ‘ 黑’ (black) + ‘犬’ (dog), which is used to mock black peo- ple

Decomposition: As Chinese characters are ideograms, they can convey specific emotions through individual characters’ separation and combination. For example, The character ‘默’(silence) is composed of ‘ 黑’ (black) + ‘犬’ (dog), which is used to mock black peo- ple. Examples: •好奇纹了什么字，默吗 I wonder what character was used to cre- ate curiosity, ‘默’?

work page