Pith. sign in

REVIEW 4 major objections 6 minor 32 references

Towards Generalizable Generic Harmful Speech Datasets for Implicit Hate Speech Detection

T0 review · 4 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read Generic hate speech data can be repurposed to train implicit hate detectors with a +12.9 F1 gain.

desk verdict The data-cleaning idea is real and the recall gains are sometimes large, but the headline F1 claim is inflated by macro-averaging and a single favorable cell; worth a serious look, not a citation yet. read the letter →

arxiv 2506.16476 v1 pith:ULZPYMRW submitted 2025-06-19 cs.CL cs.AIcs.LG

Saad Almohaimeed , Saleh Almohaimeed , Damla Turgut , Ladislau Bölöni This is my paper · ORCID
classification cs.CLcs.AIcs.LG MSC 68T50
keywords implicithatespeechdetectiondatasetreannotationinfluentialsamplesGPT-4oLlama-3dataaugmentationcross-datasetgeneralization
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

Generic harmful speech datasets (Davidson, Waseem, Founta, HateXplain) contain far more implicit hate than their labels show: a lexicon analysis by the authors finds that between 14% and 71.7% of positive samples contain no offensive words. The paper tries to establish that this latent signal can be surfaced by a three-stage pipeline—identifying influential training samples through a small trusted benchmark, correcting their labels with GPT-4o, and augmenting them with Llama-3 paraphrases—so that models trained on generic data detect implicit hate nearly as well as models trained on specialized implicit-hate datasets. If correct, the approach makes the large existing stock of generic hate datasets usable for a task that currently suffers from data scarcity, with a reported +12.9 F1 improvement over the baseline on specialized test sets and no loss on generic ones.

What carries the argument

The load-bearing mechanism is the 'influential sample' loop. A model is trained on a generic dataset; the 500-sample Trusted Samples Dataset (TSD), built by GPT-4o plus two human experts, serves as a probe; each TSD sample the model misclassifies is used to retrieve the top-x training samples whose embeddings are most cosine-similar to it and share the model's (wrong) predicted label. These retrieved samples are then processed by one of three operations—removal, GPT-4o reannotation, or Llama-3 70B augmentation—and the model is retrained in repeated loops until the best-performing version is found by manual inspection. The paper also motivates the whole approach with a lexicon-based observation that a large share of positive samples in generic datasets contain no offensive words, which it reads as evidence that implicit hate is already present but unlabeled.

What would settle it

Take the GPT-4o-reannotated versions of the influential samples and have two independent human experts re-annotate a random subset; if human agreement with the GPT-4o labels is near chance, or if training on human-corrected labels removes the +12.9 F1 gain, the cleaning step is not doing the causal work claimed. Alternatively, flip an equal number of random labels in the same datasets and retrain: if random flips reproduce the gain, the benefit comes from perturbation, not correction.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that mislabeled and confusing samples—not the general distribution of generic datasets—are what prevents implicit hate from being learned from them. The authors show that training a BERT model, finding which training samples are most responsible (by cosine similarity in the model's embedding space) for misclassifying a 500-sample trusted benchmark, and then either dropping those samples, reannotating them with GPT-4o, or appending Llama-3 paraphrased versions of them, substantially improves recall and F1 on three specialized implicit-hate test sets (IHC, OLID_IH, THOS_IH). The best configuration raises the average F1 by 12.9 points over training on the original datasets, and the improvement holds across all four generic training corpora in cross-dataset settings.

Load-bearing premise

The argument depends on GPT-4o being a trustworthy re-annotator: if it relabels genuinely harmful implicit samples as neutral, or neutral samples as harmful, the 'cleaned' training data is corrupted and the reported cross-dataset gains would not reflect real signal.

Editorial extensions

If this is right

  • Training on generic hate speech datasets after influential-sample processing can substitute for scarce specialized implicit-hate data, easing a bottleneck in hate-speech research.
  • Existing large generic corpora (Waseem, Davidson, Founta, HateXplain) become resources for implicit hate detection rather than being written off as too noisy.
  • The trusted-samples probe method gives a reusable way to find which training examples drive cross-dataset misclassifications, not just in hate speech.
  • Dropping influential samples alone already raises recall on specialized tests (e.g., Waseem recall on IHC from 0.064 to 0.640), suggesting label noise is a major blocker for generalization.
  • The pipeline's gains come without sacrificing generic performance, so a single model can cover both explicit and implicit harmful speech.

Reading between the lines

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

  • The same 'latent subclass' logic could be applied to other under-annotated categories inside generic datasets—e.g., sarcasm, disinformation, or self-harm—where the label exists implicitly but annotators did not record it.
  • The ceiling of the approach is set by the oracle quality of GPT-4o; if future LLMs surpass it on implicit-hate judgments, the same pipeline should yield further gains, and if GPT-4o's judgments are biased, the method inherits that bias.
  • A testable extension would be to replace the cosine-similarity retrieval with influence functions or gradient-based attribution and check whether the same F1 gains appear with fewer dropped samples.
  • Because the authors report that the best loop was chosen by manual inspection, an automated selection criterion could make the pipeline end-to-end and easier to reproduce across domains.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a pipeline for converting generic harmful-speech datasets into training data for implicit hate speech detection. It constructs a 500-sample Trusted Samples Dataset (TSD) with GPT-4o and two human experts, identifies 'influential' training samples by finding TSD examples the model misclassifies and retrieving the most cosine-similar training samples via the model's embeddings, and then processes those influential samples in three ways: dropping them, reannotating them with GPT-4o, or augmenting them with Llama-3 70B paraphrases. The authors train BERT on four generic datasets (Waseem, Davidson, Founta, HateXplain) and evaluate across four generic and three specialized implicit-hate datasets, reporting Recall and F1. The abstract and conclusion claim a +12.9-point F1 improvement over baseline on specialized datasets, and Section 5.2 states that all proposed approaches substantially improved performance compared to training on the original datasets.

Significance. If the claimed gains are real, the paper would make a useful practical contribution: it would repurpose abundant generic hate datasets for implicit-hate detection and show that influential-sample processing is a cheap alternative to large-scale new annotation. The cross-dataset evaluation over seven datasets is a strength, and several single-cell gains are large, e.g., Waseem + Drop Influential improves IHC recall from 0.064 to 0.640, and HateXplain reannotation/augmentation improves THOS_IH F1 from 0.462 to 0.511/0.521. However, the F1 metric is never defined and appears to be macro-averaged, the headline gain is a single cell rather than an aggregate, and the cleaning target is partly circular because GPT-4o is used both to build the TSD and as the relabeling oracle. These issues are load-bearing for the central claim that generic datasets can be converted into usable implicit-hate training data.

major comments (4)
  1. [§4.2, Table 3] The F1 score is never defined, and the reported numbers are inconsistent with binary positive-class F1. For example, Waseem→IHC baseline has Recall 0.064 and F1 0.512; with perfect precision the maximum positive-class F1 at recall 0.064 is 0.120. The F1 column must therefore be macro-averaged (or otherwise include the negative class). Macro-F1 can be dominated by the abundant negative class, so the F1 improvements in Tables 2–3 do not by themselves establish improved implicit-hate detection. Please report positive-class precision/recall/F1 (or a per-class breakdown), clarify the averaging, and add confidence intervals or significance tests for the five-run averages.
  2. [Abstract, §5.2, §6, Table 3] The headline '+12.9-point F1 improvement' is a single cell (Waseem + Drop Influential on THOS_IH: F1 0.484 vs. 0.355), not an average, and is not marked in any table. The claim in §5.2 that 'all proposed approaches substantially improved performance compared to training on the original datasets' is contradicted by Table 3: five of the 36 F1 cells are below their baseline (Davidson Drop on IHC, Founta Drop on IHC, and HateXplain Drop/Reannotation/Augmentation on IHC), and the average improvement is far smaller than 12.9 points. Please state which cells support the headline, use a clearly defined aggregate with error bars, and temper the over-general claim.
  3. [§3.2, §3.4, Limitations] There is a circularity risk in the cleaning target: GPT-4o is used to construct the TSD that identifies influential samples (§3.2), and GPT-4o is also the oracle that relabels those influential samples (§3.4). Because the 'trusted' TSD partly reflects GPT-4o's label preferences, the pipeline may tune training data toward GPT-4o's judgments rather than toward human ground truth. The Limitations section concedes that LLMs have not reached human-expert level on implicit hate. Please provide a human-agreement study on a sample of reannotated influential instances, report the fraction of label flips per dataset, and show that the improvements are robust when the oracle is validated against or replaced by human annotations on a subset.
  4. [§4.1, Limitations] The pipeline's main hyperparameters appear to be selected on the evaluation data. The number of loops is chosen by 'manual observation of the optimal loop results' (Limitations), and §4 reports results based on the best epoch. This selection on the test sets can inflate cross-dataset numbers and makes the five-seed averages difficult to interpret. Please use a held-out validation split for epoch and loop selection, or report results under a fixed rule such as the last epoch, and if manual selection is retained, quantify its effect on a separate validation set.
minor comments (6)
  1. [Table 3] The baseline that supports the abstract's +12.9 headline (Waseem + Drop Influential on THOS_IH) is not visually identifiable; please add a marker or a footnote so readers can locate the claimed gain.
  2. [References] The reference for Kim et al. (2022) contains 'COOLING-2022', which should be 'COLING-2022'.
  3. [§5.3] The word 'occured' should be 'occurred'.
  4. [§4] The learning-rate sweep is listed as '1e-6 to 1e-7'; please clarify the direction and step size of the grid, and whether the best epoch was chosen on a validation split.
  5. [§3.4, §3.5] The exact prompts used for GPT-4o reannotation and Llama-3 augmentation are not provided; including them would substantially improve reproducibility.
  6. [§3.2] The TSD is described as a contribution but no availability statement or release plan is given; please state whether the dataset will be made public.

Circularity Check

2 steps flagged · score 3.0 of 10

Central cross-dataset gains use original test labels and are not circular, but the exploration rows that reannotate test sets with GPT-4o, together with the GPT-4o-built TSD used to select influential samples, create partial self-confirming loops.

  1. self definitional [Section 5.3 / Table 2 and Table 3 last rows ('GPT4o on Train and Test'); Section 3.4]
    "For each training dataset listed in Table 2, the last row (exploration) shows results where the influential samples were reannotated in both the training and testing datasets. ... we employed GPT4o to reannotate the influential samples, and adjust the ground truth label if it does not match the GPT4o annotation."

    In the exploration rows, both the corrected training labels and the evaluation labels are produced by the same GPT-4o reannotation procedure. The reported F1 and Recall therefore measure how well the retrained model reproduces GPT-4o's label preferences, not how well it matches independent human-annotated ground truth. The improvement in those cells is partly forced by the shared oracle rather than by genuinely better implicit-hate detection. The headline +12.9 gain and most other results use original test labels, so this circularity is limited to the exploration configuration.

  2. other [Sections 3.2 and 3.4]
    "The TSD was constructed collaboratively using GPT-4o and two human experts in the field of harmful speech detection. ... GPT-4o contributed by identifying diverse targets that human experts might overlook and generating neutral samples that closely resemble harmful speech in structure and tone."

    The trusted samples used to flag influential training samples are partly generated by GPT-4o, and the same GPT-4o model is then the oracle that reannotates the flagged influential samples. Thus the criterion for 'mislabeled' and the source of the replacement labels are not independent; the loop can reinforce GPT-4o's own biases instead of grounding corrections in an external standard. Because human experts curated the TSD, this is a partial rather than total circularity, and it does not affect the original-label test results.

full rationale

The paper's central claim rests on cross-dataset evaluations using original test labels (e.g., Table 3's THOS_IH column, where the +12.9-point F1 gain appears), so the main result is not derived from the paper's own definitions or from a fitted parameter renamed as a prediction. The influential-sample selection and GPT-4o reannotation pipeline is an empirical method whose success on external benchmarks is contingent, not logically forced. However, two configurations introduce partial self-reference: the exploration rows reannotate the test sets with the same GPT-4o oracle used to correct training labels, so those improvements measure agreement with GPT-4o rather than independent human truth; and the TSD that selects influential samples is partly GPT-4o-generated while GPT-4o also supplies the relabeling. Neither loop invalidates the original-label results, but they prevent the exploration-based comparisons from being fully independent. No load-bearing self-citation chain or imported uniqueness theorem was found: the self-cited lexicon and GPT-4o-effectiveness claims are accompanied by independent references and are not the derivation engine. The macro-F1 and metric-reporting concerns raised by the reviewer are correctness risks, not circularity. Overall score 3 reflects partial, non-central circularity rather than a derivation that is equivalent to its inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 1 invented entities

The central claim rests on these assumptions: the 500-sample TSD is correct; the lexicon identifies implicit hate; GPT-4o relabeling is accurate; Llama-3 paraphrases preserve labels; and embedding cosine similarity selects the true influential samples. None of these is independently validated, and the TSD, which is the linchpin, is not released. The free parameters (top-x, loop count, best epoch) are all selected using test-set feedback, which further weakens the claim.

free parameters (3)
  • top_x_influential = 10 (Waseem, Davidson, HateXplain), 20 (Founta)
    Chosen per dataset in Section 4.1; Founta got 20 because it is larger. The value changes which training samples are flagged, so it directly controls the cleaned dataset.
  • num_loops = Waseem 16, Davidson 3, Founta 13, HateXplain 7
    Table 1; the stopping point is 'manual observation of the optimal loop results' (Section 4.1), i.e., selected on test performance rather than an independent criterion.
  • test_epoch = epoch 1-20 with best test performance per dataset
    Section 4 states all experiments run for 20 epochs and 'the reported results are based on the best epoch performance.' This chooses the model on the test set and can inflate scores.
assumptions (5)
  • domain assumption A harmful sample that contains no offensive lexicon terms is implicit hate.
    Used in Section 1 to estimate that 14%-71.7% of positive samples are implicit hate; the authors note it depends on lexicon comprehensiveness.
  • domain assumption GPT-4o labels are reliable enough to correct the training data.
    Section 3.4 replaces ground truth when GPT-4o disagrees; Limitations concede LLMs are not at human expert level for implicit hate.
  • domain assumption Llama-3 70B paraphrases of explicit hate are valid implicit-hate training samples with the same label.
    Section 3.5; the authors found 8B failed and 70B succeeded, and BERT could not converge when the augmented set was large, so augmentation was limited to influential samples.
  • domain assumption Cosine similarity in the model's embedding space identifies samples causally responsible for misclassification.
    Section 3.3 top_influence uses nearest neighbors by csim, not true influence functions; no evidence that these samples are the actual cause.
  • domain assumption The TSD's 500 labels are accurate.
    Section 3.2; TSD is treated as trusted ground truth for selecting influential samples, but its labels were produced with GPT-4o and two authors, not independently audited.
invented entities (1)
  • Trusted Samples Dataset (TSD)
    purpose: 500-sample benchmark and guide used to identify influential training samples; central to the pipeline
    No public release, no independent label audit; its construction with GPT-4o means it cannot serve as an independent oracle.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Generalizable Generic Harmful Speech Datasets for Implicit Hate Speech Detection." pith.science (2026). https://pith.science/paper/ULZPYMRW

@misc{pith2026250616476,
  author       = {Pith},
  title        = {Pith review of: Towards Generalizable Generic Harmful Speech Datasets for Implicit Hate Speech Detection},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ULZPYMRW}},
  note         = {Machine review of arXiv:2506.16476}
}
read the original abstract

Implicit hate speech has recently emerged as a critical challenge for social media platforms. While much of the research has traditionally focused on harmful speech in general, the need for generalizable techniques to detect veiled and subtle forms of hate has become increasingly pressing. Based on lexicon analysis, we hypothesize that implicit hate speech is already present in publicly available harmful speech datasets but may not have been explicitly recognized or labeled by annotators. Additionally, crowdsourced datasets are prone to mislabeling due to the complexity of the task and often influenced by annotators' subjective interpretations. In this paper, we propose an approach to address the detection of implicit hate speech and enhance generalizability across diverse datasets by leveraging existing harmful speech datasets. Our method comprises three key components: influential sample identification, reannotation, and augmentation using Llama-3 70B and GPT-4o. Experimental results demonstrate the effectiveness of our approach in improving implicit hate detection, achieving a +12.9-point F1 score improvement compared to the baseline.

Figures

Figures reproduced from arXiv: 2506.16476 by the authors.

Figure 1
Figure 1. The pipeline of our proposed methodology [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Averaged Performance Metrics for training on generic datasets and evaluated over generic and specialized [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 29 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Hyeseon Ahn, Youngwook Kim, Jungin Kim, and Yo-Sub Han. 2024. S hared C on: Implicit hate speech detection using shared semantics. In Findings of the Association for Computational Linguistics: ACL 2024. Association for Computational Linguistics

    work page 2024

  4. [4]

    Saad Almohaimeed, Saleh Almohaimeed, and Ladislau Bölöni. 2024. Transfer learning and lexicon-based approaches for implicit hate speech detection: A comparative study of human and GPT-4 annotation. In Proc. of 2024 IEEE 18th Int. Conf. on Semantic Computing (ICSC-2024), pages 142--147

    work page 2024

  5. [5]

    Saad Almohaimeed, Saleh Almohaimeed, Ashfaq Ali Shafin, Bogdan Carbunar, and Ladislau Bölöni. 2023. THOS : A benchmark dataset for targeted hate and offensive speech. In Proc. of Data-centric Machine Learning Research (DMLR) Workshop at ICML 2023

    work page 2023

  6. [6]

    Saad Almohaimeed, Saleh Almohaimeed, Damla Turgut, and Ladislau Bölöni. 2025. Closest positive cluster loss: Improving the generalization of implicit hate speech classifiers across social media datasets. In Proc. of IEEE Int. Conf. on Communications (ICC-2025)

    work page 2025

  7. [7]

    Eric Arazo, Diego Ortego, Paul Albert, Noel O’Connor, and Kevin McGuinness. 2019. Unsupervised label noise modeling and loss correction. In Proc. of Int. Conf. on Machine Learning (ICML-2019), pages 312--321

    work page 2019

  8. [8]

    Tommaso Caselli, Valerio Basile, Jelena Mitrovi \'c , Inga Kartoziya, and Michael Granitzer. 2020. I feel offended, don ' t be abusive! I mplicit/explicit messages in offensive and abusive language. In Proc. of the Twelfth Language Resources and Evaluation Conf., pages 6193--6202

    work page 2020

Show all 32 references
  1. [9]

    Thomas Davidson, Dana Warmsley, Michael Macy, and Ingmar Weber. 2017. Automated hate speech detection and the problem of offensive language. In Proc. of the Int. AAAI Conf. on Web and Social Media (ICWSM-2017), volume 11, pages 512--515

    2017

  2. [10]

    Esra D \"o nmez, Thang Vu, and Agnieszka Falenska. 2024. Please note that I ’m just an AI : A nalysis of behavior patterns of LLM s in (non-) offensive speech identification. In Proc. of the 2024 Conf. on Empirical Methods in Natural Language Processing EMNLP-2024 , pages 18340--18357

    2024

  3. [11]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The L lama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  4. [12]

    Mai ElSherief, Caleb Ziems, David Muchlinski, Vaishnavi Anupindi, Jordyn Seybolt, Munmun De Choudhury, and Diyi Yang. 2021. Latent hatred: A benchmark for understanding implicit hate speech. In Proc. of the Conf. on Empirical Methods in Natural Language Processing (EMNLP-2021)...

    2021

  5. [13]

    Antigoni Founta, Constantinos Djouvas, Despoina Chatzakou, Ilias Leontiadis, Jeremy Blackburn, Gianluca Stringhini, Athena Vakali, Michael Sirivianos, and Nicolas Kourtellis. 2018. Large scale crowdsourcing and characterization of T witter abusive behavior. In Proc. of the Int...

    2018

  6. [14]

    Xiaochuang Han and Yulia Tsvetkov. 2020. Fortifying toxic speech detectors against veiled toxicity. In Proc. of the Conf. on Empirical Methods in Natural Language Processing (EMNLP-2020), pages 7732--7739

    2020

  7. [15]

    Thomas Hartvigsen, Saadia Gabriel, Hamid Palangi, Maarten Sap, Dipankar Ray, and Ece Kamar. 2022. T oxi G en: A large-scale machine-generated dataset for adversarial and implicit hate speech detection. In Proc. of the 60th Annual Meeting of the Association for Computational Li...

    2022

  8. [16]

    Fan Huang, Haewoon Kwak, and Jisun An. 2023. Is ChatGPT better than human annotators? P otential and limitations of ChatGPT in explaining implicit hate speech. In Companion Proc. of the ACM Web Conference 2023, page 294–297

    2023

  9. [17]

    Uthman Jinadu and Yi Ding. 2024. Noise correction on subjective datasets. In Proc. of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL-2024), pages 5385--5395

    2024

  10. [18]

    Jacob Devlin Ming-Wei Chang Kenton and Lee Kristina Toutanova. 2019. BERT : Pre-training of deep bidirectional transformers for language understanding. In Proc. of the 2019 Conf. of the North American chapter of the Association for Computational Linguistics (NAACL-2019), pages...

    2019

  11. [19]

    Youngwook Kim, Shinwoo Park, and Yo-Sub Han. 2022. Generalizable implicit hate speech detection using contrastive learning. In Proc. of the 29th Int. Conf. on Computational Linguistics (COOLING-2022), pages 6667--6679

    2022

  12. [20]

    Youngwook Kim, Shinwoo Park, Youngsoo Namgoong, and Yo-Sub Han. 2023. C on P rompt: Pre-training a language model with machine-generated data for implicit hate speech detection. In Proc. of Conf. on Empirical Methods in Natural Language Processing (EMNLP 2023)

    2023

  13. [21]

    Pang Wei Koh and Percy Liang. 2017. Understanding black-box predictions via influence functions. In Proc. of Int. Conf. on Machine Learning (ICML-2017), pages 1885--1894

    2017

  14. [22]

    Sheng Liu, Jonathan Niles-Weed, Narges Razavian, and Carlos Fernandez-Granda. 2020. Early-learning regularization prevents memorization of noisy labels. Advances in Neural Information Processing Systems (NeurIPS-2020), 33:20331--20342

    2020

  15. [23]

    Binny Mathew, Punyajoy Saha, Seid Muhie Yimam, Chris Biemann, Pawan Goyal, and Animesh Mukherjee. 2021. HateXplain : A benchmark dataset for explainable hate speech detection. In Proc. of the AAAI Conf. on Artificial Intelligence (AAAI-2021), volume 35, pages 14867--14875

    2021

  16. [24]

    Nicol \'a s Benjam \' n Ocampo, Elena Cabrio, and Serena Villata. 2023. Unmasking the hidden meaning: Bridging implicit and explicit hate speech embedding representations. In Proc. of Conf. on Empirical Methods in Natural Language Processing (EMNLP 2023)

    2023

  17. [25]

    Nedjma Ousidhoum, Zizheng Lin, Hongming Zhang, Yangqiu Song, and Dit-Yan Yeung. 2019. Multilingual and multi-aspect hate speech analysis. In Proc. of the Conf. on Empirical Methods in Natural Language Processing and the 9th Int. Joint Conf. on Natural Language Processing (EMNL...

    2019

  18. [26]

    Garima Pruthi, Frederick Liu, Satyen Kale, and Mukund Sundararajan. 2020. Estimating training data influence by tracing gradient descent. Advances in Neural Information Processing Systems (NeurIPS-2020), 33:19920--19930

    2020

  19. [27]

    Scott Reed, Honglak Lee, Dragomir Anguelov, Christian Szegedy, Dumitru Erhan, and Alexander Rabinovich. 2015. Training deep neural networks on noisy labels with bootstrapping. In Proc. of Int. Conf. on Learning Representations (ICLR-2015)

    2015

  20. [28]

    Smith, and Yejin Choi

    Maarten Sap, Saadia Gabriel, Lianhui Qin, Dan Jurafsky, Noah A. Smith, and Yejin Choi. 2020. Social bias frames: Reasoning about social and power implications of language. In Proc. of the 58th Annual Meeting of the Association for Computational Linguistics (ACL-2020), pages 5477--5490

    2020

  21. [29]

    Bertie Vidgen, Tristan Thrush, Zeerak Waseem, and Douwe Kiela. 2021. Learning from the worst: Dynamically generated datasets to improve online hate detection. In Proc. of the Conf. on Empirical Methods in Natural Language Processing and the 9th Int. Joint Conf. on Natural Lang...

    2021

  22. [30]

    Zeerak Waseem and Dirk Hovy. 2016. Hateful symbols or hateful people? P redictive features for hate speech detection on T witter. In Proc. of the NAACL Student Research Workshop, pages 88--93

    2016

  23. [31]

    Marcos Zampieri, Shervin Malmasi, Preslav Nakov, Sara Rosenthal, Noura Farra, and Ritesh Kumar. 2019. Predicting the type and target of offensive posts in social media. In Proc. of the 2019 Conf. of the North American chapter of the Association for Computational Linguistics (N...

    2019

  24. [32]

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2018. mixup: Beyond empirical risk minimization. In Proc. of Int. Conf. on Learning Representations (ICLR-2018)

    2018

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.