Pith. sign in

REVIEW 4 major objections 5 minor 77 references

Detoxify: A framework for abusive text transformation using LLMs

T0 review · 4 major / 5 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 The paper finds that LLM detoxification quality depends on the model: Groq over-positivizes and alters meaning, while GPT-4o and DeepSeek preserve it.

desk verdict The headline result is a confound: Groq is hardware running Llama-3-8B, so 'Groq stood out' really means an 8B model behaved differently than frontier models. read the letter →

arxiv 2507.10177 v2 pith:FUZKT2LU submitted 2025-07-14 cs.CL cs.AI

Rohitash Chandra , Jiyong Choi , Jayesh Sonawane This is my paper · ORCID
classification cs.CLcs.AI
keywords abusivetexttransformationlargelanguagemodelshatespeechdetectionsentimentanalysissemanticsimilaritydetoxificationLLMevaluationsocialmediamoderation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper proposes Detoxify, a framework that uses large language models to rewrite abusive tweets and reviews into polite text while keeping the original message. It evaluates four models—Gemini, GPT-4o, DeepSeek, and Groq—on abuse detection and transformation, then scores the rewrites with sentiment analysis, keyword counts, n-gram patterns, and semantic similarity. The central finding is that Groq behaves distinctly: it adds the most positive phrasing, produces the fewest successful transformations, and yields the lowest semantic similarity to the input, while GPT-4o and DeepSeek cluster together as the most faithful rewrites. A sympathetic reader would read this as evidence that the choice of LLM materially changes whether automated moderation preserves a speaker's intent.

What carries the argument

The framework's load-bearing machinery is a six-stage evaluation pipeline: data cleaning, LLM API configuration, abuse detection, transformation, and then three independent scores—SenWave-BERT sentiment classification, keyword counting with a log-odds ratio using two abusive-word datasets, and MPNet cosine similarity. The n-gram (bigram/trigram) analysis is what visually separates the models: Groq and Gemini insert new positive phrases, while GPT-4o and DeepSeek retain original content words like 'sharia law'. The semantic similarity tables (original versus each model, and model versus model) carry the main comparative claim about Groq's distinctness.

What would settle it

Re-run the 400-tweet transformation with every model under identical safety-filter settings—disabling Gemini's filters as before and matching Groq's constraints as closely as the APIs allow—then recompute cosine similarity to the original; if Groq's distinctness disappears, the paper's central behavioral claim does not survive.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that LLM-based detoxification is not a homogeneous operation: the four models divide into two behaviors. Groq consistently restructures abusive text into longer, conspicuously positive language—its top trigram after transformation is 'let work together'—to the point where the original context is sometimes lost or altered, and its transformed tweets have the lowest average cosine similarity to the originals. Gemini also leans positive but rewrites less extensively, whereas GPT-4o and DeepSeek preserve the original n-grams and phrasing most closely, sharing the highest pairwise semantic similarity of any pair. The paper's conclusion therefore states that Groq stood out as the most distinct model, with excessive positive phrasing that occasionally cost it the original meaning.

Load-bearing premise

The comparison assumes the four models' output differences reflect intrinsic behavior, but Gemini was run with all safety filters disabled while Groq could not disable its filters, so the observed gap could shrink if settings were equalized.

Editorial extensions

If this is right

  • If Groq's rewrites drift semantically, moderation systems that use it must add a fidelity check or risk silently changing what a user meant.
  • GPT-4o and DeepSeek's near-identical behavior suggests either model can serve as a detoxification default when preserving original content is the priority.
  • The framework's combination of n-gram, sentiment, and semantic scoring can be applied to any new LLM without retraining the evaluation models.
  • Because all models reduced hate-word counts sharply, the paper implies detoxification-by-LLM is feasible, with model choice trading off fidelity versus positivity.

Reading between the lines

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

  • One untested implication is that the reported Groq distinctness may be partly an artifact of API safety configuration rather than the underlying model: Gemini's filters were off, while Groq's were not adjustable.
  • A practical extension would be a fidelity-versus-positivity benchmark that scores rewrites on both toxicity reduction and semantic preservation, letting platforms tune which model to deploy.
  • The similarity of GPT-4o and DeepSeek also raises a cost question the paper leaves open: if a cheaper or open-weight model matches a frontier model on this task, the cheaper one may be sufficient for real moderation pipelines.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Detoxify, a pipeline that uses LLMs to rewrite abusive tweets and reviews into polite text while preserving sentiment and semantics. It compares Gemini-1.5-Flash, GPT-4o, DeepSeek-V3, and 'Groq' (identified in Section 3.4 as llama3-8b-8192) on abuse detection and transformation, and evaluates outputs with n-gram, keyword, sentiment (SenWave-BERT), and semantic (MPNet cosine similarity) analyses. The headline finding is that Groq was the most distinct: it produced the most positive rephrasings, the lowest transformation success, and the lowest semantic similarity to the source. The paper also reports that GPT-4o and DeepSeek behaved similarly.

Significance. If the comparison were valid, the paper would provide a useful practical benchmark of LLM-based text detoxification: it covers two datasets, releases code and data, and combines several evaluation lenses (sentiment, semantics, n-grams, keyword counts). However, the central comparison is undermined by a model-identity confound (an 8B open-weight model is labelled as a 'state-of-the-art LLM' and compared with much larger proprietary models) and an API-safety confound (Gemini had all safety filters disabled while Groq did not). The abuse-detection results are also difficult to interpret because the Stage 4 prompt appears to be the transformation prompt. These issues bear directly on the paper's only headline claims, so the current version cannot support its conclusions.

major comments (4)
  1. [Sections 2.2, 3.4, and 6] Groq is a hardware platform, not an LLM; the system actually evaluated is llama3-8b-8192. Comparing it with Gemini-1.5-Flash, GPT-4o, and DeepSeek-V3 mixes model identity and scale into the reported 'distinctness' of Groq. The conclusion in Section 6 ('Groq stood out as the most distinct') is therefore not supported as a statement about Groq versus other LLMs. A matched control (the same model served by different providers, or different model sizes on the same provider) would be needed.
  2. [Section 3.4 and Section 4.2] The prompt labelled for Stage 4 abuse detection is identical to the Stage 5 transformation prompt and contains no instruction to classify or count abusive words. It is therefore unclear how the binary accuracy, precision, recall, F1, and IoU values in Table 2 and Section 4.2 were derived; they may be measuring transformation outputs. The authors should state explicitly how abusive/non-abusive labels were produced from a prompt that asks only for the transformed tweet.
  3. [Sections 4.3 and 4.5, Tables 3 and 4] The term 'transformation success rate' is never defined. The paper reports large differences (Gemini 53.4% vs Groq 18.4% in Table 3; GPT-4o 396/400 vs Groq 371/400 in Table 4) but does not state what counts as a success: no rubric, automatic check, human annotation, or instruction-following criterion is given. Without this definition the central quantitative comparisons cannot be reproduced or interpreted.
  4. [Section 4.3] Gemini was run with HARM_BLOCK_THRESHOLD set to BLOCK NONE (all safety filters disabled), whereas Groq 'did not offer such flexibility in safety settings' and GPT-4o/DeepSeek had no configurable safety options. This API-configuration difference confounds transformation success, sentiment shift, verbosity, and semantic similarity. The paper acknowledges the effect on success rates but still attributes Groq's distinct behavior to the model in Section 6; equalizing safety settings would be necessary to make the comparison meaningful.
minor comments (5)
  1. [Abstract and Section 1] The Abstract and Section 1 say 'two state-of-the-art LLMs' while the study actually evaluates four; the Abstract also writes 'DeekSeek' instead of DeepSeek.
  2. [Sections 3.3 and 3.4] Section 3.3 states the framework has six stages but then enumerates seven stages (Stages 1-7); additionally, the Stage 4 and Stage 5 prompts in Section 3.4 are identical, which obscures the intended distinction between detection and transformation.
  3. [Section 4.2, Tables 2 and 3] Table 2's caption calls the results 'sentiment accuracy', but Section 4.2 describes the task as abusive/non-abusive classification; Section 4.2 also refers to Groq's accuracy as shown in Table 3, although Table 3 is the transformation table in Section 4.3. Captions and cross-references should be corrected.
  4. [Section 4.1 and Figure 2] The terms 'clean reviews' and 'clean improved' in Figure 2 and Section 4.1 are not defined; it is unclear which model produced the 'improved' trigrams and how the displayed examples were selected.
  5. [Section 4.6] Section 4.6 reports that HateBERT classified every tweet as hateful and that a keyword-search method was used instead; the keyword list and the details of the 'log-odds ratio with informative Dirichlet priors' computation should be described, and the limitation should be stated more prominently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the central Groq-distinctness claim is a direct empirical observation, not a derived quantity.

full rationale

The paper contains no derivation that reduces to its own inputs. The headline result—that Groq 'stood out as the most distinct'—is a reporting of observed metrics: transformation success counts (Table 4), n-gram frequencies (Figure 10), SenWave-BERT sentiment distributions (Figure 13), and MPNet cosine similarities (Tables 6-7). These are measured, not fitted, and the qualitative conclusion is a summary of those measurements rather than an output of a model tuned on them. The safety-setting asymmetry (Gemini run with BLOCK NONE while Groq 'did not offer such flexibility in safety settings') and the fact that 'Groq' is a hardware platform serving llama3-8b-8192 are genuine threats to the comparison's validity, but they are confounds, not circular steps: no equation defines the conclusion in terms of its own assumptions. Self-citations (e.g., [23], [34], [51], [60], [68]) appear in background and tool-choice contexts; none supplies a uniqueness theorem or a fitted parameter that later reappears as the paper's result, so they do not make the argument circular. The study is self-contained as an empirical comparison against external datasets and pre-trained evaluation models.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The paper introduces no mathematical derivation and no new entities. The central claim rests on measurement assumptions: that the sentiment and semantic evaluators are valid for transformed text, that LLM API outputs are treated as reproducible model behavior, and that the keyword lists capture residual hate. The hand-chosen API prompt and safety settings also act as free configuration choices.

free parameters (1)
  • LLM API prompt and safety settings = Tailored prompt; Gemini HARM_BLOCK_THRESHOLD=4 (BLOCK NONE)
    The paper reports tuning API configurations and prompts to the task (Stage 3), and Gemini was run with safety disabled while Groq was not, so the comparison depends on hand-chosen settings.
assumptions (3)
  • domain assumption The BERT-based sentiment model (SenWave-BERT) provides a valid measure of sentiment for both raw and transformed tweets.
    The paper relies on SenWave-BERT for all sentiment comparisons in Section 4.7.1 without validating it on transformed text.
  • domain assumption LLM API outputs are treated as reproducible model behavior, despite non-deterministic sampling and unspecified temperature settings.
    No temperature or seed settings are reported for any of the four LLM APIs in Section 3.4.
  • domain assumption The keyword lists derived from the SWAD and Dynamically-Generated-Hate-Speech datasets are sufficient to measure residual hate in transformed text.
    The keyword-search evaluation in Section 4.6 replaces failed HateBERT analysis but no threshold or log-odds criteria are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detoxify: A framework for abusive text transformation using LLMs." pith.science (2026). https://pith.science/paper/FUZKT2LU

@misc{pith2026250710177,
  author       = {Pith},
  title        = {Pith review of: Detoxify: A framework for abusive text transformation using LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FUZKT2LU}},
  note         = {Machine review of arXiv:2507.10177}
}
read the original abstract

Although Large Language Models (LLMs) have demonstrated significant advancements in natural language processing tasks, their effectiveness in the classification and transformation of abusive text into non-abusive versions remains an area for exploration. In this study, we present Detoxify: a framework that employs LLMs to transform abusive text (tweets and reviews) containing hate speech and profanity into non-abusive text while retaining the original intent. We evaluate the performance of four state-of-the-art LLMs, such as Gemini, GPT-4o, DeekSeek and Groq, on their ability to identify abusive text. We aim to transform and obtain a text that is clean of abusive and inappropriate content, but maintains a similar level of sentiment and semantics, i.e. the transformed text needs to maintain its message. Afterwards, we evaluate the raw and transformed datasets with sentiment analysis and semantic analysis. Our results show Groq provides vastly different results when compared with other LLMs. We have identified similarities between GPT-4o and DeepSeek. Groq stood out as the most distinct, as it often restructured sentences with excessive positive phrasing, with the original context lost or altered.

Figures

Figures reproduced from arXiv: 2507.10177 by the authors.

Figure 1
Figure 1. Framework Diagram This early evaluation highlighted the strengths and limitations of each model, helping us identify areas where further fine￾tuning could enhance performance for the larger dataset. These preliminary results laid the groundwork for the subsequent stages of our framework. We also perform bigram and trigram analysis to identify the most common sequence of words, then compare the transformation rates b… view at source ↗
Figure 2
Figure 2. Trigram comparison using Groq for showing original and transformed trigrams for the IIT dataset [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Batch-wise sentiment accuracy comparison of Groq and Gemini [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (12 more)
Figure 4
Figure 4. Figure 4: Batch-wise transformation accuracy comparison of Groq and Gemini [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Count of hate after keyword search by model [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Count of hate by model and type of abuse [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Line plot of hate count by batch [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Successful transformation counts by Model and Abuse type [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Bigram and Trigram by category 9 [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: Bigram and Trigram comparisons by model 10 [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 12
Figure 12. Figure 12: Heatmap of sentiment by model type [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]
Figure 13
Figure 13. Figure 13: shows that prior to any transformations, the most common sentiment by far was ”annoyed”, followed by ”jok￾ing”. For each of the transformations, we see a dramatic de￾crease in ”annoyed” tweets across all four models and a mas￾sive increase in the ”optimistic” category…
Figure 11
Figure 11. Figure 11: , 12 shows that the most prevalent categories are ”annoyed”, ”optimistic”, ”joking” and ”official report”. The remaining categories are all below 100, including only 3 thank￾ful categories. There were no classifications of ”empathy” or ”pessimism”, likely due to the L…
Figure 15
Figure 15. Figure 15: Semantic analysis visualisations: pairwise similarity and UMAP [PITH_FULL_IMAGE:figures/full_fig_p014_15.png]
Figure 14
Figure 14. Figure 14: Cosine similarity of semantic scores by batch [PITH_FULL_IMAGE:figures/full_fig_p014_14.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

77 extracted references · 47 canonical work pages

  1. [10]

    Olkhovikov, G

    G. Olkhovikov, G. Badia, Maximality of bi-intuitionistic propositional logic, Journal of Logic and Computation 32 (1) (2022) 1–31

    work page 2022

  2. [26]

    Sridhar, C

    D. Sridhar, C. De Bacco, D. Blei, Estimating social influence from obser- vational data, in: Conference on Causal Learning and Reasoning, PMLR, 2022, pp. 712–733

    work page 2022

  3. [48]

    Guarcello, P

    C. Guarcello, P. Solinas, A. Braggio, M. Di Ventra, F. Giazotto, Josephson thermal memory, Physical Review Applied 9 (1) (2018) 014021

    work page 2018

  4. [1]

    C. T. Carr, R. A. Hayes, Social media: Defining, developing, and divining, Atlantic journal of communication 23 (1) (2015) 46–65

    work page 2015

  5. [2]

    A. M. Kaplan, M. Haenlein, Users of the world, unite! the challenges and opportunities of social media, Business horizons 53 (1) (2010) 59–68

    work page 2010

  6. [3]

    Mitsu, E

    R. Mitsu, E. Dawood, Cyberbullying: An overview, Indonesian Journal of Global Health Research 4 (2022) 195–202. doi:10.37287/ijghr. v4i1.927

    work page doi:10.37287/ijghr 2022

  7. [4]

    J. A. Pater, M. K. Kim, E. D. Mynatt, C. Fiesler, Characterizations of online harassment: Comparing policies across social media platforms, in: Proceedings of the 2016 ACM International Conference on Supporting Group Work, 2016, pp. 369–374

    work page 2016

  8. [5]

    Van Laer, The means to justify the end: Combating cyber harassment in social media, Journal of Business Ethics 123 (1) (2014) 85–98

    T. Van Laer, The means to justify the end: Combating cyber harassment in social media, Journal of Business Ethics 123 (1) (2014) 85–98

    work page 2014

Show all 77 references
  1. [6]

    Y . Wu, X. Wang, S. Hong, M. Hong, M. Pei, Y . Su, The relationship between social short-form videos and youth’s well-being: It depends on usage types and content categories., Psychology of Popular Media 10 (4) (2021) 467

    2021

  2. [7]

    Hudders, S

    L. Hudders, S. De Jans, M. De Veirman, The commercialization of social media stars: a literature review and conceptual framework on the strate- gic use of social media influencers, Social media influencers in strategic communication (2021) 24–67

    2021

  3. [8]

    Richards, P

    D. Richards, P. H. Caldwell, H. Go, Impact of social media on the health of children and young people, Journal of paediatrics and child health 51 (12) (2015) 1152–1157

    2015

  4. [9]

    Zampieri, P

    M. Zampieri, P. Nakov, S. Rosenthal, P. Atanasova, G. Karadzhov, H. Mubarak, L. Derczynski, Z. Pitenis, C ¸ . C ¸¨oltekin, Semeval-2020 task 12: Multilingual offensive language identification in social media (offen- seval 2020), arXiv preprint arXiv:2006.07235 (2020)

    2020 arXiv

  5. [11]

    Cinelli, G

    M. Cinelli, G. D. F. Morales, A. Galeazzi, W. Quattrociocchi, M. Starnini, The echo chamber e ffect on social media, Proceedings of the National Academy of Sciences 118 (9) (2021) e2023301118. doi:10.1073/ pnas.2023301118

    2021

  6. [12]

    V . U. Gongane, M. V . Munot, A. D. Anuse, Detection and moderation of detrimental content on social media platforms: current status and future directions, Social Network Analysis and Mining 12 (1) (2022) 129

    2022

  7. [13]

    Myers West, Censored, suspended, shadowbanned: User interpreta- tions of content moderation on social media platforms, New Media & Society 20 (11) (2018) 4366–4383

    S. Myers West, Censored, suspended, shadowbanned: User interpreta- tions of content moderation on social media platforms, New Media & Society 20 (11) (2018) 4366–4383

    2018

  8. [14]

    C. D. Manning, H. Sch ¨utze, Foundations of Statistical Natural Language Processing, MIT Press, Cambridge, MA, 1999

    1999

  9. [15]

    M. S. Jahan, M. Oussalah, A systematic review of hate speech automatic detection using natural language processing, Neurocomputing 546 (2023) 126232

    2023

  10. [16]

    Alrashidi, A

    B. Alrashidi, A. Jamal, I. Khan, A. Alkhathlan, A review on abusive con- tent automatic detection: approaches, challenges and opportunities, PeerJ Computer Science 8 (2022) e1142

    2022

  11. [17]

    Schmidt, M

    A. Schmidt, M. Wiegand, A survey on hate speech detection using natural language processing, in: Proceedings of the fifth international workshop on natural language processing for social media, 2017, pp. 1–10

    2017

  12. [18]

    D. W. Otter, J. R. Medina, J. K. Kalita, A survey of the usages of deep learning for natural language processing, IEEE transactions on neural net- works and learning systems 32 (2) (2020) 604–624

    2020

  13. [19]

    Torfi, R

    A. Torfi, R. A. Shirvani, Y . Keneshloo, N. Tavaf, E. A. Fox, Natural language processing advancements by deep learning: A survey, arXiv preprint arXiv:2003.01200 (2020)

    2020 arXiv

  14. [20]

    Wankhade, A

    M. Wankhade, A. C. S. Rao, C. Kulkarni, A survey on sentiment analy- sis methods, applications, and challenges, Artificial Intelligence Review 55 (7) (2022) 5731–5780

    2022

  15. [21]

    T. B. Brown, Language models are few-shot learners, arXiv preprint arXiv:2005.14165 (2020)

    2020 arXiv

  16. [22]

    Naveed, A

    H. Naveed, A. U. Khan, S. Qiu, M. Saqib, S. Anwar, M. Usman, N. Akhtar, N. Barnes, A. Mian, A comprehensive overview of large lan- guage models (2024). arXiv:2307.06435. URL https://arxiv.org/abs/2307.06435

    2024 arXiv

  17. [23]

    Chandra, A

    R. Chandra, A. Krishna, Covid-19 sentiment analysis via deep learning during the rise of novel cases, PLOS ONE 16 (2021) 1–26. doi:10. 1371/journal.pone.0255615. URL https://doi.org/10.1371/journal.pone.0255615

    2021 doi

  18. [24]

    H. Dong, S. Xie, Large language models (llms): Deployment, tokenomics and sustainability, arXiv preprint arXiv:2405.17147 (2024)

    2024 arXiv

  19. [25]

    G. Team, T. Mesnard, C. Hardin, et al., Gemma: Open models based on Gemini research and technology, arXiv preprint arXiv:2403.08295 (2024)

    2024 arXiv

  20. [27]

    Bandi, P

    A. Bandi, P. V . S. R. Adapa, Y . E. V . P. K. Kuchi, The power of generative ai: A review of requirements, models, input–output formats, evaluation metrics, and challenges, Future Internet 15 (8) (2023) 260

    2023

  21. [28]

    Gozalo-Brizuela, E

    R. Gozalo-Brizuela, E. C. Garrido-Merchan, Chatgpt is not all you need. a state of the art review of large generative ai models, arXiv preprint arXiv:2301.04655 (2023)

    2023 arXiv

  22. [29]

    Amirifar, S

    T. Amirifar, S. Lahmiri, M. K. Zanjani, An nlp-deep learning approach for product rating prediction based on online reviews and product features, IEEE Transactions on Computational Social Systems (2023)

    2023

  23. [30]

    Jaremko, D

    J. Jaremko, D. Gromann, M. Wiegand, Revisiting implicitly abusive lan- guage detection: Evaluating llms in zero-shot and few-shot settings, in: Proceedings of the 31st International Conference on Computational Lin- guistics, 2025, pp. 3879–3898

    2025

  24. [31]

    O. Buruk, Academic writing with gpt-3.5 (chatgpt): reflections on prac- tices, efficacy and transparency, in: Proceedings of the 26th International Academic Mindtrek Conference, 2023, pp. 144–153

    2023

  25. [32]

    Megawati, H

    R. Megawati, H. Listiani, N. W. Pranoto, M. Akobiarek, et al., The role of gpt chat in writing scientific articles: A systematic literature review, Jurnal Penelitian Pendidikan IPA 9 (11) (2023) 1078–1084

    2023

  26. [33]

    Lazebnik, A

    T. Lazebnik, A. Rosenfeld, Detecting llm-assisted writing in scientific communication: Are we there yet?, Science 9 (3) (2024) 1–10

    2024

  27. [34]

    Chandra, A

    R. Chandra, A. Chaudhary, Y . Rayavarapu, An evaluation of LLMs and Google Translate for translation of selected indian languages via senti- ment and semantic analyses, arXiv preprint arXiv:2503.21393 (2025)

    2025 arXiv

  28. [35]

    Novgorodov, I

    S. Novgorodov, I. Guy, G. Elad, K. Radinsky, Generating product de- scriptions from user reviews, in: The world wide web conference, 2019, pp. 1354–1364

    2019

  29. [36]

    D. Rothman, Transformers for Natural Language Processing: Build, train, and fine-tune deep neural network architectures for NLP with Python, Hugging Face, and OpenAI’s GPT-3, ChatGPT, and GPT-4, Packt Pub- lishing Ltd, 2022

    2022

  30. [37]

    O’Shea, R

    K. O’Shea, R. Nash, An introduction to convolutional neural networks (2015). arXiv:1511.08458. URL https://arxiv.org/abs/1511.08458

    2015 arXiv

  31. [38]

    Graves, J

    A. Graves, J. Schmidhuber, Framewise phoneme classification with bidi- rectional lstm and other neural network architectures, Neural Networks 18 (5) (2005) 602–610, iJCNN 2005. doi:https://doi.org/10. 1016/j.neunet.2005.06.042. URL https://www.sciencedirect.com/science/article...

    2005

  32. [39]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need (2023). arXiv:1706. 03762. URL https://arxiv.org/abs/1706.03762

    2023 arXiv

  33. [40]

    X. Li, Q. Yu, B. Alzahrani, A. Barnawi, A. Alhindi, D. Alghazzawi, Y . Miao, Data fusion for intelligent crowd monitoring and management systems: A survey, IEEE Access 9 (2021) 47069–47083

    2021

  34. [41]

    Zhang, D

    Z. Zhang, D. Robinson, J. Tepper, Detecting hate speech on twitter using a convolution-gru based deep neural network, in: The Semantic Web: 15th International Conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018, Proceedings 15, Springer, 2018, pp. 745–760

    2018

  35. [42]

    Devlin, BERT: Pre-training of deep bidirectional transformers for lan- guage understanding, arXiv preprint arXiv:1810.04805 (2018)

    J. Devlin, BERT: Pre-training of deep bidirectional transformers for lan- guage understanding, arXiv preprint arXiv:1810.04805 (2018)

    2018 arXiv

  36. [43]

    A. A. Hind Saleh, K. Moria, Detection of hate speech using bert and hate speech word embedding with deep model, Applied Arti- ficial Intelligence 37 (1) (2023) 2166719. arXiv:https://doi. org/10.1080/08839514.2023.2166719, doi:10.1080/08839514. 2023.2166719. URL https://doi.or...

    2023

  37. [44]

    Mishra, H

    P. Mishra, H. Yannakoudakis, E. Shutova, Tackling online abuse: A survey of automated abuse detection methods, arXiv preprint arXiv:1908.06024 (2019)

    2019 arXiv

  38. [45]

    C. Yan, S. Wang, J. Yang, T. Xu, J. Huang, Re-balancing variational au- 17 toencoder loss for molecule sequence generation, in: Proceedings of the 11th ACM international conference on bioinformatics, computational bi- ology and health informatics, 2020, pp. 1–7

    2020

  39. [46]

    M. Sap, D. Card, S. Gabriel, Y . Choi, N. A. Smith, The risk of racial bias in hate speech detection, in: Proceedings of the 57th annual meeting of the association for computational linguistics, 2019, pp. 1668–1678

    2019

  40. [47]

    Dixon, J

    L. Dixon, J. Li, J. Sorensen, N. Thain, L. Vasserman, Measuring and mitigating unintended bias in text classification, in: Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society, 2018, pp. 67– 73

    2018

  41. [49]

    Rasmy, Y

    L. Rasmy, Y . Xiang, Z. Xie, C. Tao, D. Zhi, Med-bert: pre-trained con- textualized embeddings on large-scale structured electronic health records for disease prediction (2020). arXiv:2005.12833. URL https://arxiv.org/abs/2005.12833

    2020 arXiv

  42. [50]

    Caselli, V

    T. Caselli, V . Basile, J. Mitrovi ´c, M. Granitzer, HateBERT: Retraining BERT for abusive language detection in English, in: A. Mostafazadeh Da- vani, D. Kiela, M. Lambert, B. Vidgen, V . Prabhakaran, Z. Waseem (Eds.), Proceedings of the 5th Workshop on Online Abuse and Harms...

    2021

  43. [51]

    URL https://aclanthology.org/2021.woah-1.3/

    doi:10.18653/v1/2021.woah-1.3. URL https://aclanthology.org/2021.woah-1.3/

    2021 doi

  44. [52]

    Chandra, G

    R. Chandra, G. Ren, Group-H, Longitudinal abuse and sentiment analysis of hollywood movie dialogues using llms (2025). arXiv:2501.13948. URL https://arxiv.org/abs/2501.13948

    2025

  45. [53]

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, et al., Gemini: A family of highly capable multimodal models (2024). arXiv:2312.11805. URL https://arxiv.org/abs/2312.11805

    2024 arXiv

  46. [54]

    G. Team, P. Georgiev, V . I. Lei, R. Burnell, L. Bai, A. Gulati, G. Tanzer, D. Vincent, Z. Pan, S. Wang, S. Mariooryad, et al., Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context (2024). arXiv:2403.05530. URL https://arxiv.org/abs/2403.05530

    2024 arXiv

  47. [55]

    Grattafiori, A

    A. Grattafiori, A. Dubey, A. Jauhri, A. Pandey, A. Kadian, A. Al-Dahle, et al., The Llama 3 herd of models (2024). arXiv:2407.21783. URL https://arxiv.org/abs/2407.21783

    2024 arXiv

  48. [56]

    Achiam, S

    OpenAI, J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, a. p. u. others. year=2024, eprint=2303.08774, Gpt-4 technical report

    2024 arXiv

  49. [57]

    DeepSeek-AI, D. Guo, D. Yang, H. Zhang, J. Song, R. Zhang, R. Xu, Q. Zhu, S. Ma, P. Wang, X. Bi, et al., Deepseek-r1: Incentivizing rea- soning capability in llms via reinforcement learning (2025). arXiv: 2501.12948. URL https://arxiv.org/abs/2501.12948

    2025 arXiv

  50. [58]

    DeepSeek-AI, A. Liu, B. Feng, B. Xue, B. Wang, B. Wu, C. Lu, C. Zhao, C. Deng, C. Zhang, C. Ruan, D. Dai, D. Guo, D. Yang, D. Chen, D. Ji, E. Li, F. Lin, et al., Deepseek-v3 technical report (2025). arXiv:2412. 19437. URL https://arxiv.org/abs/2412.19437

    2025 arXiv

  51. [60]

    User Profiling and Abusive Language Detection Dataset, UCI Machine Learning Repository, DOI: https://doi.org/10.24432/C57K66 (2019)

    2019 doi

  52. [61]

    Chandra, V

    R. Chandra, V . Kulkarni, Semantic and sentiment analysis of se- lected bhagavad gita translations using bert-based language framework, IEEE Access 10 (2022) 21291–21315. doi:10.1109/ACCESS.2022. 3152266

    2022 doi

  53. [62]

    Davidson, D

    T. Davidson, D. Warmsley, M. Macy, I. Weber, Automated hate speech detection and the problem of o ffensive language, in: Proceedings of the international AAAI conference on web and social media, V ol. 11, 2017, pp. 512–515

    2017

  54. [63]

    Q. Yang, H. Alamro, S. Albaradei, A. Salhi, X. Lv, C. Ma, M. Alshehri, I. Jaber, F. Tifratene, W. Wang, T. Gojobori, C. M. Duarte, X. Gao, X. Zhang, Senwave: Monitoring the global sentiments under the covid-19 pandemic (2020). arXiv:2006.10842. URL https://arxiv.org/abs/2006.10842

    2020 arXiv

  55. [64]

    Mikolov, K

    T. Mikolov, K. Chen, G. Corrado, J. Dean, E fficient estimation of word representations in vector space (2013). arXiv:1301.3781. URL https://arxiv.org/abs/1301.3781

    2013 arXiv

  56. [65]

    D. P. Kingma, Adam: A method for stochastic optimization, arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  57. [66]

    E. W. Pamungkas, V . Basile, V . Patti, Do You Really Want to Hurt Me? predicting abusive swearing in social media, in: Proceedings of the Twelfth Language Resources and Evaluation Conference, European Lan- guage Resources Association, Marseille, France, 2020, pp. 6237–6246. U...

    2020

  58. [67]

    Vidgen, T

    B. Vidgen, T. Thrush, Z. Waseem, D. Kiela, Learning from the worst: Dynamically generated datasets to improve online hate detection, in: C. Zong, F. Xia, W. Li, R. Navigli (Eds.), Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 1...

    2021 doi

  59. [68]

    J. Park, J. Yang, A. Tolbert, K. Bunsold, You change the way you talk: Examining the network, toxicity and discourse of cross-platform users on twitter and parler during the 2020 us presidential election, Journal of Information Science (04 2024). doi:10.1177/01655515241238405

    2020 doi

  60. [69]

    Singh, R

    A. Singh, R. Chandra, Hp-bert: A fine-tuned bert model for detecting hinduphobia and sentiment analysis (2025). doi:10.5281/zenodo. 14898138. URL https://doi.org/10.5281/zenodo.14898138

    2025 doi

  61. [70]

    K. Song, X. Tan, T. Qin, J. Lu, T. Liu, Mpnet: Masked and permuted pre-training for language understanding, CoRR abs /2004.09297 (2020). arXiv:2004.09297. URL https://arxiv.org/abs/2004.09297

    2020 arXiv

  62. [71]

    McInnes, J

    L. McInnes, J. Healy, J. Melville, Umap: Uniform manifold approxima- tion and projection for dimension reduction (2020).arXiv:1802.03426. URL https://arxiv.org/abs/1802.03426

    2020 arXiv

  63. [72]

    Barbieri, J

    F. Barbieri, J. Camacho-Collados, F. Ronzano, L. Espinosa-Anke, M. Ballesteros, V . Basile, V . Patti, H. Saggion, SemEval 2018 task 2: Multilingual emoji prediction, in: M. Apidianaki, S. M. Mohammad, J. May, E. Shutova, S. Bethard, M. Carpuat (Eds.), Proceedings of the 12th ...

    2018 doi

  64. [73]

    Sykora, S

    M. Sykora, S. Elayan, T. W. Jackson, A qualitative analysis of sar- casm, irony and related #hashtags on twitter, Big Data & Society 7 (2) (2020) 2053951720972735. arXiv:https://doi.org/10. 1177/2053951720972735, doi:10.1177/2053951720972735. URL https://doi.org/10.1177/205395...

    2020 doi

  65. [74]

    Founta, C

    A. Founta, C. Djouvas, D. Chatzakou, I. Leontiadis, J. Blackburn, G. Stringhini, A. Vakali, M. Sirivianos, N. Kourtellis, Large scale crowd- sourcing and characterization of twitter abusive behavior, Proceedings of the International AAAI Conference on Web and Social Media 12 (...

    2018 doi

  66. [75]

    Walia, Toxicity in the digital age: A review of literature on online toxic behaviors and social media platforms (05 2023)

    B. Walia, Toxicity in the digital age: A review of literature on online toxic behaviors and social media platforms (05 2023). doi:10.13140/RG.2. 2.15541.08165

    2023

  67. [76]

    Fortuna, S

    P. Fortuna, S. Nunes, A survey on automatic detection of hate speech in text, ACM Comput. Surv. 51 (4) (Jul. 2018). doi:10.1145/3232676. URL https://doi.org/10.1145/3232676

    2018 doi

  68. [77]

    A. Q. Jiang, A. Sablayrolles, A. Mensch, C. Bamford, D. S. Chap- lot, D. de las Casas, F. Bressand, G. Lengyel, G. Lample, L. Saulnier, L. R. Lavaud, M.-A. Lachaux, P. Stock, T. L. Scao, T. Lavril, T. Wang, T. Lacroix, W. E. Sayed, Mistral 7b (2023). arXiv:2310.06825. URL http...

    2023 arXiv

  69. [78]

    Ngiam, A

    J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, A. Y . Ng, Multimodal deep learning, in: Proceedings of the 28th International Conference on Interna- tional Conference on Machine Learning, ICML’11, Omnipress, Madison, WI, USA, 2011, p. 689–696. 18

    2011

Pith tools

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