Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

Detecting harassment and defamation in cyberbullying with emotion-adaptive training

T0 review · 4 major / 4 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Training transformer models on emotion-labeled text before fine-tuning them on a few hundred cyberbullying examples raises average macro F1, precision, and recall by about 20 percent across nine architectures.

desk verdict The few-shot EAT result is real and reproducible, but the defamation class is a genre confound, so the headline 20% claim is not yet interpretable as cyberbullying detection. read the letter →

arxiv 2501.16925 v1 pith:W46BRKDV submitted 2025-01-28 cs.CL

Peiling Yi , Arkaitz Zubiaga , Yunfei Long This is my paper · ORCID
classification cs.CL
keywords cyberbullyingdetectiondefamationharassmentemotion-adaptivetrainingdomainadaptationlow-resourcetextclassificationtransformermodelscelebrity
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

This paper tries to establish that a simple two-stage training recipe—first teach a transformer model to recognize emotions, then fine-tune it on a small set of cyberbullying examples—substantially improves detection of both harassment and defamation. The authors build a three-class celebrity cyberbullying dataset, HDCyberbullying, and test nine transformer architectures in low-resource settings. They report that their emotion-adaptive training (EAT) raises average macro F1, precision, and recall by about 20 percent across all nine models. The result matters because indirect cyberbullying such as defamation is common and poorly served by existing binary harassment detectors.

What carries the argument

The mechanism is the EAT framework, a three-phase domain-adaptation pipeline. First, source domain selection picks a large, single-genre emotion dataset (58k Reddit comments labeled with 27 emotions) and checks its overlap with the cyberbullying domain using PCA-reduced embeddings. Second, domain concept shift groups and maps emotion labels onto the three cyberbullying classes—anger and disgust become harassment, surprise becomes defamation, joy and gratitude become non-cyberbullying—so that the conditional distributions align. Third, knowledge transfer adapts a pretrained transformer on the emotion data (zero-shot, unsupervised) and then continues training on a few cyberbullying examples (few-shot, semi-supervised).

What would settle it

Re-annotate the 250 defamation instances from scratch with multiple independent annotators using the paper's own definition of celebrity cyberbullying; if agreement with the inherited labels is low, or if the articles are judged to be news rather than social-media comments, then the 20 percent improvement is a text-genre artifact. Alternatively, replace the defamation class with short social-media comments labeled for defamation and check whether EAT still lifts F1.

Watch

Extended reading notes

Core claim

The paper's central claim is that EAT turns a weak low-resource cyberbullying classifier into a strong one by transferring knowledge from emotion detection. In the authors' experiments, standard fine-tuning on 10 percent of the data gives near-zero F1 for the defamation class; after first training on emotion-labeled text and then fine-tuning on the same small set, nine models consistently reach macro F1 around 0.78 on average, with defamation F1 rising from 0 to values between roughly 0.5 and 0.76. They attribute this to a domain shift: emotions such as anger and disgust map to harassment, surprise ties to defamation, and joy and gratitude mark non-cyberbullying, so the emotion task teaches the model distinctions that the scarce cyberbullying labels cannot.

Load-bearing premise

The defamation examples are long celebrity news articles drawn from a fake-news dataset with labels inherited and 'manually verified' without reported inter-annotator agreement; if those articles are not actually cyberbullying, the reported improvement measures genre classification rather than defamation detection.

Editorial extensions

If this is right

  • Across the nine tested transformer architectures, EAT raises average macro F1, precision, and recall by about 20 percent relative to fine-tuning on the small cyberbullying set alone.
  • The defamation class, which baseline fine-tuning misses entirely (F1 of 0 for masked models such as RoBERTa, BERT, DistilBERT, MPNet, and ELECTRA), becomes detectable after EAT.
  • EAT is most valuable when data is scarcest: with roughly 210 training examples it matches the performance of the baseline trained on about 1,300 examples.
  • The benefit holds across six different pretraining approaches, including masked, autoregressive, replaced-token, text-to-text, and decoder-only LLMs, so the recipe is not tied to one architecture.

Reading between the lines

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

  • A natural next experiment is to test EAT on social-media-native defamation labels rather than the long fake-news articles used here; if the gain survives, the method becomes a general low-resource recipe for indirect abuse detection.
  • The authors' own error analysis shows that 70 percent of defamation samples are predicted as joy or gratitude, which suggests that positive-affect indicators, not just negative ones, may carry the signal; a testable variant would remap approval or admiration to defamation and measure the effect.
  • The reported domain similarity (average cosine similarity 0.993) may partly reflect the long article length shared by the two corpora; a matched-length control would test whether the transfer signal is semantic rather than stylistic.
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 / 4 minor

Summary. The paper introduces HDCyberbullying, a dataset of celebrity-directed harassment and defamation text (harassment from social-media comments, defamation inherited from a fake-news dataset of full articles), and proposes an emotion-adaptive training (EAT) framework that first fine-tunes transformer models on the GoEmotions emotion dataset (with an emotion-to-class mapping) and then fine-tunes on a small amount of target cyberbullying data. The authors evaluate nine transformer-based models under zero-shot and few-shot settings, reporting that EAT improves average macro F1, precision, and recall by 20% in cyberbullying detection under low-resource settings. They also provide error analyses, t-SNE visualizations, and an analysis of emotion likelihoods.

Significance. If the central claim holds, the paper would make a useful contribution: it addresses an underexplored indirect-cyberbullying setting, proposes a simple transfer-learning recipe from emotion to cyberbullying detection, and evaluates across a wide range of model architectures. The paper's strengths include the public release of code/data, the multi-model comparison, the few-shot learning-curve analysis, and the qualitative error analysis that provides insight into model misclassifications. However, as detailed below, the validity of the evaluation benchmark and the scope of the empirical claims need substantial clarification before the results can be accepted.

major comments (4)
  1. [Results (Table 3)] The headline claim that 'EAT consistently improves the average macro F1, precision and recall by 20%' is not supported for the zero-shot configuration. Table 3 shows that zero-shot EAT does not improve over the baseline overall (macro F1 0.56 vs 0.57; macro precision 0.58 vs 0.61), while only the few-shot configuration shows the large gains (macro F1 0.78 vs 0.57). The abstract and conclusion state the improvement without this scope restriction. Please revise the claim to specify few-shot EAT, and report the exact computation of the '20%' (absolute percentage points or relative improvement) so that the reader can verify it from Table 3.
  2. [HDCyberbullying (Selection of comments, Annotation, Data statistics)] The defamation class in HDCyberbullying is sourced from the Perez-Rosas et al. fake-news dataset, which consists of full celebrity news articles (average length 1,985 words) rather than social-media comments (harassment average length 455 words). Because defamation accounts for only 8% of instances but receives one-third of the macro-F1 weight, the large few-shot improvements on this class (e.g., RoBERTa's defamation F1 rising from 0 to 0.76) could reflect the model learning to distinguish long news articles from short comments, rather than detecting defamatory cyberbullying. The manual verification of inherited labels is described without reporting inter-annotator agreement or demonstrating that these articles constitute 'defamation' as defined by the paper ('speaking half-truths or lies') rather than simply 'fake news.' This issue is load-bearing for the central claim: if the benchmark is not actually measuring cyberbullying, the reported improvement is not interpretable. Please provide additional validation of the defamation labels, or report results with a re-annotated or re-collected defamation sample.
  3. [Domain concept shift (Figure 4, Figure 10)] The emotion-to-class mapping (Anger/Disgust to Harassment, Surprise to Defamation, Gratitude/Joy to Non-cyberbullying) is a central design choice of EAT, but the paper provides no quantitative justification or ablation for this mapping. The paper's own likelihood analysis (Figure 10) shows that the chosen emotions are not the maximum-likelihood emotions for their assigned classes: 'Neutral' has the highest likelihood for all three classes, and 'approval' is more likely than 'surprise' for defamation. Since the transfer mechanism is entirely mediated by this mapping, the observed few-shot improvements may depend on the particular mapping rather than on a general emotion-cyberbullying relationship. Please include an ablation over alternative emotion mappings (or a principled selection procedure), and discuss how the results change.
  4. [Evaluation (Table 3, Settings)] The metric definitions are unclear. The table header labels the F1 column 'Micro F1' while the row 'A' is described as 'Macro-averaging,' and the text says the paper reports per-class recall, precision, and micro-F1 plus macro-averaged values. It is not clear which metric is used in the 'Overall' row and whether the 20% improvement is based on macro-averaged F1 across classes and models. Please clarify the exact metric definitions and ensure the reported improvement is computed consistently and reproducibly.
minor comments (4)
  1. [Training loss (Figure 7)] The x-axis of Figure 7 is labeled only as 'size'; please specify that it represents the number of emotion training samples, and define the loss curve being plotted.
  2. [Insights from EAT (Emotion reference)] The sentence 'In conclusion, reducing source classifier training loss is general, but need to consider more factors when a lack of target samples is present' is grammatically incomplete; please rephrase for clarity.
  3. [Theoretical Analysis] The theoretical analysis is informal and does not connect the cited transfer-learning theory (Ben-David et al. 2010; Yosinski et al. 2014) to any specific bound or condition under which EAT should improve performance. Please state the assumptions and the intended logical role of this section more explicitly.
  4. [Table 4] In Table 4, the column meanings are ambiguous (e.g., the 'Emotion', 'Cyberbullying', and 'Prediction' columns for sample 3 appear to mix gold labels and model predictions). Please clarify the table header and legend.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: EAT's held-out evaluation does not reduce to the emotion mapping or to the paper's own inputs.

full rationale

The paper's central claim is an empirical comparison: models are first trained on GoEmotions labels mapped to three cyberbullying-related classes, then fine-tuned on a small held-out split of HDCyberbullying and tested on the remaining 90%. The test set is not used to define the training objective, and no equation in the paper makes the reported macro-F1 a tautology of the emotion mapping. The emotion-to-class mapping (Anger/Disgust -> Harassment, Surprise -> Defamation, Gratitude/Joy -> Non-bullying) is an input design choice, not a fitted parameter. In fact, the paper's own Figure 10 shows that the chosen emotions are not the maximum-likelihood emotions for the target classes (Neutral is most effective, and Approval is more aligned with defamation than Surprise), which cuts against the idea that the mapping was selected to force the reported results. The phrase "preliminary experiments" is vague, but the paper does not exhibit any test-set fitting, and all reported numbers are averages over five runs on held-out splits. Self-citations to Yi and Zubiaga (2022, 2023a, 2023b) are contextual background and are not load-bearing for the EAT derivation. The concern that the defamation class reuses fake-news articles and may measure text-genre classification is a construct-validity issue about the benchmark, not a circularity in the derivation chain: the F1 score is not equivalent to the training input by construction. No circular step meeting the required evidence standard was found.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The framework rests on four assumptions: domain relatedness, the emotion mapping, dataset label validity, and the domain-similarity measure. No numeric free parameters are fitted to the test labels, but the emotion mapping and source dataset size are experimenter choices that affect the headline result.

free parameters (1)
  • Emotion source dataset size = 3700 (varied in Figure 7)
    Number of GoEmotions samples used for emotion training; chosen by the experimenters rather than derived. Figure 7 shows the relationship between source data size, training loss, and downstream performance is non-monotonic, so the choice affects the reported gains.
assumptions (4)
  • domain assumption The emotion detection domain and the cyberbullying detection domain are related enough that a data space in GoEmotions can transfer knowledge to cyberbullying.
    Stated in Methodology, Problem definition: 'We assume that the emotion detection domain and the cyberbullying detection domain are related but from different distributions.' This is the load-bearing premise of EAT.
  • ad hoc to paper The emotion-to-class mapping (Anger/Disgust to Harassment, Surprise to Defamation, Gratitude/Joy to Non-cyberbullying) is a valid bridge between the source and target label spaces.
    Defined in 'Domain concept shift'. The choice is justified by prior research and 'preliminary experiments', but the paper's own Figure 10 shows Neutral is the most likely emotion for all target classes and Approval is more associated with defamation than Surprise, so the mapping is not independently established.
  • domain assumption The HDCyberbullying labels are trustworthy: combining pre-existing harassment and fake-news labels with manual verification yields correct harassment, defamation, and non-cyberbullying categories.
    Assumed in 'Annotation'. No inter-annotator agreement or detailed annotation protocol is reported; a single manual verification pass is described.
  • standard math PCA on RoBERTa embeddings and an average cosine similarity of 0.993 indicates high domain similarity between emotion and cyberbullying data.
    Used in 'Source domain selection' to justify the source choice. The interpretation assumes the reduced 2-D cosine similarity is a meaningful measure of domain overlap, which is not independently validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Detecting harassment and defamation in cyberbullying with emotion-adaptive training." pith.science (2026). https://pith.science/paper/W46BRKDV

@misc{pith2026250116925,
  author       = {Pith},
  title        = {Pith review of: Detecting harassment and defamation in cyberbullying with emotion-adaptive training},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/W46BRKDV}},
  note         = {Machine review of arXiv:2501.16925}
}
read the original abstract

Existing research on detecting cyberbullying incidents on social media has primarily concentrated on harassment and is typically approached as a binary classification task. However, cyberbullying encompasses various forms, such as denigration and harassment, which celebrities frequently face. Furthermore, suitable training data for these diverse forms of cyberbullying remains scarce. In this study, we first develop a celebrity cyberbullying dataset that encompasses two distinct types of incidents: harassment and defamation. We investigate various types of transformer-based models, namely masked (RoBERTa, Bert and DistilBert), replacing(Electra), autoregressive (XLnet), masked&permuted (Mpnet), text-text (T5) and large language models (Llama2 and Llama3) under low source settings. We find that they perform competitively on explicit harassment binary detection. However, their performance is substantially lower on harassment and denigration multi-classification tasks. Therefore, we propose an emotion-adaptive training framework (EAT) that helps transfer knowledge from the domain of emotion detection to the domain of cyberbullying detection to help detect indirect cyberbullying events. EAT consistently improves the average macro F1, precision and recall by 20% in cyberbullying detection tasks across nine transformer-based models under low-resource settings. Our claims are supported by intuitive theoretical insights and extensive experiments.

Figures

Figures reproduced from arXiv: 2501.16925 by the authors.

Figure 1
Figure 1. Examples of celebrity cyberbullying, with key [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Architecture of EAT In this section, we detail our simple yet effective domain adaptation approach, which consists of three main phases: source domain selection, domain concept shift and knowl￾edge transfer. The framework is depicted in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Domain similarity. Orange: Emotion data; Blue: [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Mapping between Emotion and Cyberbullying do [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Performance of EAT in the defamation detection [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Confusion matrix of EAT. 0: No-cyberbullying; 1: [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Training loss. The x-axis represents the size of the [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 9
Figure 9. Figure 9: Capacity of transfer. The x-axis represents the size [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]
Figure 10
Figure 10. Figure 10: The likelihood of a certain emotion identifying [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Cyberbullying Detection via Aggression-Enhanced Prompting

    cs.CL 2025-08 reject novelty 4.0 of 10

    Adding predicted aggression labels to prompts improves LLM cyberbullying detection F1 on a single dataset, but the claim is confounded by prompt-format changes and missing statistics.

Reference graph

Works this paper leans on

60 extracted references · 51 canonical work pages · cited by 1 Pith paper

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.a...

  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]

    Agrawal, S.; and Awekar, A. 2018 a . Deep learning for detecting cyberbullying across multiple social media platforms . Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), 10772 LNCS(Table 2): 141--153

    work page 2018

  4. [4]

    Agrawal, S.; and Awekar, A. 2018 b . Deep learning for detecting cyberbullying across multiple social media platforms. In European conference on information retrieval, 141--153. Springer

    work page 2018

  5. [5]

    Al-Hashedi, M.; Soon, L.-K.; Goh, H.-N.; Lim, A. H. L.; and Siew, E.-G. 2023. Cyberbullying detection based on emotion. IEEE Access

    work page 2023

  6. [6]

    AlKhamissi, B.; Li, M.; Celikyilmaz, A.; Diab, M.; and Ghazvininejad, M. 2022. A review on language models as knowledge bases. arXiv preprint arXiv:2204.06031

    arXiv 2022

  7. [7]

    Ben-David, S.; Blitzer, J.; Crammer, K.; Kulesza, A.; Pereira, F.; and Vaughan, J. W. 2010. A theory of learning from different domains . Machine Learning, 79(1-2): 151--175

    work page 2010

  8. [8]

    Bostan, L.-A.-M.; and Klinger, R. 2018. An Analysis of Annotated Corpora for Emotion Classification in Text. In Bender, E. M.; Derczynski, L.; and Isabelle, P., eds., Proceedings of the 27th International Conference on Computational Linguistics, 2104--2119. Santa Fe, New Mexico, USA: Association for Computational Linguistics

    work page 2018

Show all 60 references
  1. [9]

    Buechel, S.; and Hahn, U. 2022. Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. arXiv preprint arXiv:2205.01996

    2022 arXiv

  2. [10]

    W.; and Hinduja, S

    Burgess-Proctor, A.; Patchin, J. W.; and Hinduja, S. 2009. Cyberbullying and online harassment: Reconceptualizing the victimization of adolescent girls. Female crime victims: Reality reconsidered, 162: 176

    2009

  3. [11]

    S.; and Liu, H

    Cheng, L.; Guo, R.; Candan, K. S.; and Liu, H. 2020. Representation learning for imbalanced cross-domain classification . Proceedings of the 2020 SIAM International Conference on Data Mining, SDM 2020, 478--486

    2020

  4. [12]

    V.; and Manning, C

    Clark, K.; Luong, M.-T.; Le, Q. V.; and Manning, C. D. 2020. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. In International Conference on Learning Representations, 1--18

    2020

  5. [13]

    Dadvar, M.; and Eckert, K. 2018. Cyberbullying detection in social networks using deep learning based models . In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), volume 12393 LNCS, 245--255. ...

    2018

  6. [14]

    Demszky, D.; Movshovitz-Attias, D.; Ko, J.; Cowen, A.; Nemade, G.; and Ravi, S. 2020. G o E motions: A Dataset of Fine-Grained Emotions. In Jurafsky, D.; Chai, J.; Schluter, N.; and Tetreault, J., eds., Proceedings of the 58th Annual Meeting of the Association for Computationa...

    2020

  7. [15]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tec...

    2019

  8. [16]

    Ekman, P. 1992. Are there basic emotions?

    1992

  9. [17]

    Gururangan, S.; Marasovi \'c , A.; Swayamdipta, S.; Lo, K.; Beltagy, I.; Downey, D.; and Smith, N. A. 2020. Don’t Stop Pretraining: Adapt Language Models to Domains and Tasks. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, 8342--8360

    2020

  10. [18]

    Kajiwara, T.; Chu, C.; Takemura, N.; Nakashima, Y.; and Nagahara, H. 2021. WRIME: A new dataset for emotional intensity estimation with subjective and objective annotations. In Proceedings of the 2021 conference of the North American chapter of the association for computationa...

    2021

  11. [19]

    Karthika, C. 2022. Cyberbullying on celebrities: A case study on actress Parvathy. In AIP Conference Proceedings, volume 2463. AIP Publishing

    2022

  12. [20]

    J.; and De Choudhury, M

    Kim, S.; Razi, A.; Stringhini, G.; Wisniewski, P. J.; and De Choudhury, M. 2021. A human-centered systematic literature review of cyberbullying detection algorithms. Proceedings of the ACM on Human-Computer Interaction, 5(CSCW2): 1--34

    2021

  13. [21]

    Li, L.; Fan, L.; Atreja, S.; and Hemphill, L. 2023. “HOT” ChatGPT: The promise of ChatGPT in detecting and discriminating hateful, offensive, and toxic comments on social media. ACM Transactions on the Web

    2023

  14. [22]

    Liu, C.; Osama, M.; and De Andrade, A. 2019. DENS: A dataset for multi-class emotion analysis. arXiv preprint arXiv:1910.11769

    2019 arXiv

  15. [23]

    Liu, P.; Yuan, W.; Fu, J.; Jiang, Z.; Hayashi, H.; and Neubig, G. 2023. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1--35

    2023

  16. [24]

    Liu, Y.; Ott, M.; Goyal, N.; Du, J.; Joshi, M.; Chen, D.; Levy, O.; Lewis, M.; Zettlemoyer, L.; and Stoyanov, V. 2019. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692

    2019 arXiv

  17. [25]

    Liu, Z.; Zhang, T.; Yang, K.; Thompson, P.; Yu, Z.; and Ananiadou, S. 2024. Emotion detection for misinformation: A review. Information Fusion, 102300

    2024

  18. [26]

    H.; Laghi, F.; Baiocco, R.; Pallini, S.; and Brunner, T

    Lonigro, A.; Schneider, B. H.; Laghi, F.; Baiocco, R.; Pallini, S.; and Brunner, T. 2015. Is cyberbullying related to trait or state anger? Child Psychiatry & Human Development, 46: 445--454

    2015

  19. [27]

    Maftei, A.; Holman, A.-C.; and Merlici, I.-A. 2022. Using fake news as means of cyber-bullying: The link with compulsive internet use and online moral disengagement. Computers in Human Behavior, 127: 107032

    2022

  20. [28]

    Muneer, A.; and Fati, S. M. 2020. A comparative analysis of machine learning techniques for cyberbullying detection on twitter. Future Internet, 12(11): 187

    2020

  21. [29]

    Oberl \"a nder, L. A. M.; Kim, E.; and Klinger, R. 2020. GoodNewsEveryone: A corpus of news headlines annotated with emotions, semantic roles, and reader perception. In Proceedings of the Twelfth Language Resources and Evaluation Conference, 1554--1566

    2020

  22. [30]

    J.; and Vandebosch, H

    Ouvrein, G.; De Backer, C. J.; and Vandebosch, H. 2018. Online celebrity aggression: A combination of low empathy and high moral disengagement? The relationship between empathy and moral disengagement and adolescents' online celebrity aggression. Computers in Human Behavior, 8...

    2018

  23. [31]

    Ouyang, L.; Wu, J.; Jiang, X.; Almeida, D.; Wainwright, C.; Mishkin, P.; Zhang, C.; Agarwal, S.; Slama, K.; Ray, A.; et al. 2022. Training language models to follow instructions with human feedback. Advances in neural information processing systems, 35: 27730--27744

    2022

  24. [32]

    Peled, Y. 2019. Cyberbullying and its influence on academic, social, and emotional development of undergraduate students. Heliyon, 5(3)

    2019

  25. [33]

    P \'e rez-Rosas, V.; Kleinberg, B.; Lefevre, A.; and Mihalcea, R. 2018. Automatic Detection of Fake News. In Proceedings of the 27th International Conference on Computational Linguistics, 3391--3401

    2018

  26. [34]

    Plutchik, R. 1980. A general psychoevolutionary theory of emotion. In Theories of emotion, 3--33. Elsevier

    1980

  27. [35]

    Qiu, J.; and Golman, R. 2024. Curiosity in news consumption. Applied Cognitive Psychology, 38(2): e4195

    2024

  28. [36]

    Raffel, C.; Shazeer, N.; Roberts, A.; Lee, K.; Narang, S.; Matena, M.; Zhou, Y.; Li, W.; and Liu, P. J. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140): 1--67

    2020

  29. [37]

    Rajapakse, T. C. 2019. Simple Transformers. https://github.com/ThilinaRajapakse/simpletransformers

    2019

  30. [38]

    Saengprang, S.; and Gadavanij, S. 2021. Cyberbullying: The Case of Public Figures. LEARN Journal: Language Education and Acquisition Research Network, 14(1): 344--369

    2021

  31. [39]

    R.; and Bhatia, M

    Sangwan, S. R.; and Bhatia, M. 2020. Denigration bullying resolution using wolf search optimized online reputation rumour detection. Procedia computer science, 173: 305--314

    2020

  32. [40]

    R.; and Bhatia, M

    Sangwan, S. R.; and Bhatia, M. 2022. D-BullyRumbler: a safety rumble strip to resolve online denigration bullying using a hybrid filter-wrapper approach. Multimedia Systems, 28(6): 1987--2003

    2022

  33. [41]

    Sanh, V.; Debut, L.; Chaumond, J.; and Wolf, T. 2019. DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108

    2019 arXiv

  34. [42]

    Scheithauer, H.; Schultze-Krumbholz, A.; Pfetsch, J.; and Hess, M. 2021. Types of cyberbullying. The Wiley Blackwell handbook of bullying: A comprehensive and international review of research and intervention, 1: 120--138

    2021

  35. [43]

    R.; and Wallbott, H

    Scherer, K. R.; and Wallbott, H. G. 1994. Evidence for universality and cultural variation of differential emotion response patterning. Journal of personality and social psychology, 66(2): 310

    1994

  36. [44]

    K.; Mahdavi, J.; Carvalho, M.; Fisher, S.; Russell, S.; and Tippett, N

    Smith, P. K.; Mahdavi, J.; Carvalho, M.; Fisher, S.; Russell, S.; and Tippett, N. 2008. Cyberbullying: Its nature and impact in secondary school pupils . Journal of Child Psychology and Psychiatry and Allied Disciplines, 49(4): 376--385

    2008

  37. [45]

    Song, K.; Tan, X.; Qin, T.; Lu, J.; and Liu, T.-Y. 2020. Mpnet: Masked and permuted pre-training for language understanding. Advances in Neural Information Processing Systems, 33: 16857--16867

    2020

  38. [46]

    Stanford. 2022. StanfordNERTagger4.2.0 . https://nlp.stanford.edu/software/. [Online; accessed 16-April-2024]

    2022

  39. [47]

    Sun, C.; Qiu, X.; Xu, Y.; and Huang, X. 2020. How to Fine-Tune BERT for Text Classification? In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), volume 11856 LNAI, 194--206. Springer. ISBN 978...

    2020

  40. [48]

    surekharamireddy. 2021. Malignant Comment Classification . https://www.kaggle.com/datasets/surekharamireddy/malignant-comment-classification/. [Online; accessed 16-April-2024]

    2021

  41. [49]

    Takano, M.; Taka, F.; Ogiue, C.; and Nagata, N. 2022. Online Harassment of Celebrities and Influencers. arXiv preprint arXiv:2210.07599

    2022 arXiv

  42. [50]

    Takano, M.; Taka, F.; Ogiue, C.; and Nagata, N. 2024. Online Harassment of Japanese Celebrities and Influencers. Frontiers in Psychology, 15: 1386146

    2024

  43. [51]

    Tan, W.-K.; and Hsu, C. Y. 2023. The application of emotions, sharing motivations, and psychological distance in examining the intention to share COVID-19-related fake news. Online Information Review, 47(1): 59--80

    2023

  44. [52]

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

    2023 arXiv

  45. [53]

    Vogel, E. A. 2021. The State of Online Harassment

    2021

  46. [54]

    Wang, X.; Yang, L.; Yang, J.; Wang, P.; and Lei, L. 2017. Trait anger and cyberbullying among young adults: A moderated mediation model of moral disengagement and moral identity. Computers in Human Behavior, 73: 519--526

    2017

  47. [55]

    R.; and Le, Q

    Yang, Z.; Dai, Z.; Yang, Y.; Carbonell, J.; Salakhutdinov, R. R.; and Le, Q. V. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. Advances in neural information processing systems, 32

    2019

  48. [56]

    Yi, P.; and Zubiaga, A. 2022. Cyberbullying detection across social media platforms via platform-aware adversarial encoding. In Proceedings of the International AAAI Conference on Web and Social Media, volume 16, 1430--1434

    2022

  49. [57]

    Yi, P.; and Zubiaga, A. 2023 a . Learning like human annotators: Cyberbullying detection in lengthy social media sessions. In Proceedings of the ACM Web Conference 2023, 4095--4103

    2023

  50. [58]

    Yi, P.; and Zubiaga, A. 2023 b . Session-based cyberbullying detection in social media: A survey. Online Social Networks and Media, 36: 100250

    2023

  51. [59]

    Yosinski, J.; Clune, J.; Bengio, Y.; and Lipson, H. 2014. How transferable are features in deep neural networks? In Advances in Neural Information Processing Systems, volume 4, 3320--3328

    2014

  52. [60]

    S.; Alzain, M

    Yuvaraj, N.; Srihari, K.; Dhiman, G.; Somasundaram, K.; Sharma, A.; Rajeskannan, S.; Soni, M.; Gaba, G. S.; Alzain, M. A.; and Masud, M. 2021. Nature-Inspired-Based Approach for Automated Cyberbullying Classification on Multimedia Social Networking . Mathematical Problems in E...

    2021

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.