Pith. sign in

REVIEW 4 major objections 5 minor 33 references

Context-Aware Content Moderation for German Newspaper Comments

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Article title, topic path, and a user's moderation ratio lift LSTM and CNN models close to transformer-level performance on German comment moderation, while the same context leaves GPT-3.5 Turbo zero-shot accuracy flat.

desk verdict Useful German moderation study, but the user-ratio feature has a temporal leakage problem that inflates the reported gains. read the letter →

arxiv 2505.20963 v1 pith:V4WMYX4H submitted 2025-05-27 cs.CL cs.AI

Felix Krejca , Tobias Kietreiber , Alexander Buchelt , Sebastian Neumaier This is my paper
classification cs.CLcs.AI
keywords AutomaticContentModerationContext-AwareTextClassificationHateSpeechDetectionGermannewspapercommentsOneMillionPostsCorpusLSTMCNNLargeLanguageModels
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, for binary content moderation of German newspaper comments, small recurrent (LSTM) and convolutional (CNN) networks can be pushed close to transformer-level accuracy by adding platform-specific context. The context signals are the article's title, its topic path, and an online ratio computed from the user's own past comments, concatenated into the model input under special tokens. On the One Million Posts Corpus, the best model--an LSTM taking the comment, title, path, and full online ratio--reaches 73.3% accuracy and 0.809 AUROC (area under the ROC curve), competitive with published BERT results on related German forum data. In contrast, GPT-3.5 Turbo with seven prompt variants scores between 0.606 and 0.635 accuracy and does not improve when context is added. If the finding holds, newsrooms can get most of the benefit of large transformers from cheaper context-aware LSTM/CNN models, and should not assume that richer prompts help an LLM moderate better.

What carries the argument

The load-bearing piece is the engineered scalar $R^f_o$, the full online ratio: the fraction of a user's comments, counted over the training set plus the part of the corpus lost to downsampling, that remained online. The simpler ratio $R^s_o$ uses only the training set; both exclude test and validation. This scalar compresses the user's moderation history into one number, and the ordering in Table 6 shows that configurations carrying it are the top performers, while the path variable contributes little. Around that scalar, the model concatenates the comment, article title, and topic path as separate fields bounded by special tokens, so a single forward pass can weigh what was said against where and under which article it was said.

What would settle it

Train the best model twice on the same corpus, once with $R^f_o$ computed from the real user labels and once with $R^f_o$ computed from labels randomly permuted across users; if the two AUROC values are nearly identical, the reported context benefit is not coming from the online-ratio feature, and if the real-ratio model stays around 0.808 while the permuted model drops, then the user-history signal is genuinely carrying the gain.

Watch

Extended reading notes

Core claim

The paper's central discovery is that context carries the performance, not the model class. Among the self-trained models, every configuration that adds the article title and topic path beats the comment-only LSTM (accuracy 0.661, AUROC 0.728), and the largest jumps come from adding the full online ratio $R^f_o$: the best model, adv_LSTM_Title_path_ratio, reaches 0.733 accuracy, 0.809 AUROC, and 0.713 F1. On the GPT side, the base zero-shot prompt reaches 0.632 accuracy and the context-enriched prompts do no better, with some variants dropping to 0.606 and all producing some missing-answer refusals. The paper contrasts these numbers with the published BERT result (AUROC 0.914 on a different dataset) and concludes that context-augmented CNN/LSTM models are competitive with transformer-based approaches while being cheaper to train and deploy.

Load-bearing premise

The online-ratio feature $R^f_o$ is built from actual moderation outcomes attached to the user's comments, and the paper assumes this label-derived statistic is legitimate context rather than an indirect leak that carries the target through the user identity.

Editorial extensions

If this is right

  • A newsroom without the compute for fine-tuned transformers can deploy a context-aware LSTM/CNN and stay near published BERT-level performance on newspaper comment moderation.
  • The online-ratio feature works because past moderation outcomes of a user are predictive of future ones, so platforms with rich user histories have a real advantage in automated moderation.
  • For GPT-3.5 Turbo pipelines, the simplest prompt is the rational default, since added context raises token cost and missing-answer rate without improving accuracy.
  • The balanced precision and recall of the advanced LSTM/CNN models makes them usable as a pre-filter that sends only uncertain cases to human moderators, directly addressing moderator workload.

Reading between the lines

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

  • Because a deployment system would need the online ratio to be available before a new comment is moderated, evaluating with temporal or author-level splits is the natural next step for this line of work.
  • The same context recipe (title, path, user-history ratio) transfers naturally to other moderated forums, provided the platform logs deletions; on platforms without a public deletion history the ratio would have to be re-derived from moderator flags.
  • The negative result for GPT-3.5 Turbo should not be read as a statement about all LLMs, since the paper itself notes that newer models may use context differently; a natural successor experiment is the same prompt grid on a more recent model.
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 presents binary content-moderation classifiers for German newspaper comments using the One Million Posts Corpus. It compares LSTM, CNN, and GPT-3.5 Turbo models with and without contextual inputs (article title, topic path, and user online-ratio features). The authors report that context-aware LSTM/CNN models reach an accuracy of 0.733 and AUROC of 0.809, outperform their context-free counterparts and all GPT-3.5 zero-shot variants, and claim competitiveness with BERT-based prior work on German newspaper moderation.

Significance. If the reported results hold, the contribution is a practical, resource-efficient alternative to transformer-based moderation for German-language forums, and the study of which contextual signals help (and do not help) is useful. The paper uses a large, realistic corpus and compares several model families, including a zero-shot LLM. The finding that GPT-3.5-Turbo does not benefit from added context and underperforms smaller self-trained models is interesting and worth reporting. However, the validity of the central claim is currently compromised by a temporal leakage issue in the user-ratio feature and by the absence of a same-data transformer baseline, so the quantitative conclusions need to be re-established under a clean evaluation protocol.

major comments (4)
  1. [§3.1, Eq. (1)] The feature R^f_o is defined as (C^train_online + C^ds_online) / (C^train + C^ds), and the train/test split is random rather than temporal. For a given test comment, the user's comments posted after that comment can appear in the training set or in the downsampled set, so this feature can encode future moderation outcomes that would not be available at decision time in deployment. The note that C^ds excludes test/validation instances prevents direct leakage of the target comment's label, but it does not prevent future-comment leakage. Since the best-performing models in Table 6 (adv_LSTM_Title_path_ratio, adv_LSTM_Title_ratio, and all adv_CNN variants) all include R^f_o, the reported improvement from base_LSTM (0.661) to adv_LSTM_Title_path_ratio (0.733) may be substantially inflated by this leakage. Please re-run the experiments with a temporal split (e.g., train on comments before a cutoff date, validate and test on later comments) and recompute R^s_o and R^f_o using only comments strictly preceding each target comment; in addition, report results for all models without any user-history feature.
  2. [Table 6] No error bars, confidence intervals, or significance tests are reported anywhere in the results section. Differences among the top models (accuracy 0.733, 0.729, 0.728, 0.726, 0.726, 0.725) are small and could easily arise from random variation, especially without multiple runs. The claim that advanced LSTM and CNN models 'consistently outperform' simpler approaches, and the statement that 'there are no significant differences in performance between the advanced LSTM and CNN architectures,' are not statistically supported. Please report multiple random seeds with standard deviations and perform paired significance tests (e.g., McNemar's test on the same test set).
  3. [§4, "Comparison with Previous Literature"] The headline claim that the authors' LSTM/CNN models 'achieve competitive performance compared to state-of-the-art transformer-based models such as BERT' is not supported by the presented evidence. No transformer model is fine-tuned or evaluated on the same data, split, and label protocol used here; the comparisons with Assenmacher et al. [2], Yadav & Milde [31], Pachinger et al. [21], and Keller et al. [12] involve different datasets, different annotation guidelines, and different label distributions. Differences in AUROC and F1 can therefore reflect data rather than method. Please add a same-data BERT/GermanBERT baseline trained on the identical preprocessing and split, or explicitly weaken the claim to 'comparable to published numbers on other datasets.'
  4. [§3.1 and Table 6] The R^f_o feature is directly computed from moderation outcome labels of the user's other comments, making it a quasi-outcome feature. Even after fixing the temporal leakage, the paper should justify that, in a deployed moderation system, the moderator would have access to the historical online/offline statuses of the user's previous comments, and should ablate this feature to show how much of the reported context gain comes from title/path alone as opposed to user-outcome history. Currently the incremental contribution of R^f_o is confounded with the leakage issue, and the statement in Section 5 that 'user posting history' is a legitimate context variable needs an explicit causal-timing argument.
minor comments (5)
  1. [§3.1, text after Eq. (1)] The prose describing R^s_o and R^f_o is reversed: R^s_o is computed from the training set only, while R^f_o is computed from training plus the downsampled set. The current text says the opposite and should be corrected.
  2. [Figure 1 and Table 4] Figure 1 uses the names 'adv CNN 1 title ratio' and 'adv CNN 2 title ratio', while Table 4 uses 'adv_CNN_1_title_ratio' and 'adv_CNN_2_title_ratio'. These naming inconsistencies should be harmonized throughout the paper.
  3. [References] Reference [20], cited as the OpenAI API, points to a Mistral announcement URL rather than the OpenAI API documentation. The URL should be corrected.
  4. [Table 6] The 'Missing Answer' column for GPT models is not explained in the evaluation section. Please state explicitly how missing answers were handled when computing accuracy and F1-score.
  5. [§3.1] Please report the exact sizes of the train, validation, and test splits and the random seeds used for downsampling and splitting. This information is important for reproducibility and for assessing the temporal leakage concern.

Circularity Check

1 steps flagged · score 6.0 of 10

The R^f_o 'context' feature is computed from the target moderation labels, so the main reported gains are partly self-predictive label leakage.

  1. self definitional [Section 3.1 (Data Overview and Preprocessing), definition of R^s_o and R^f_o; best-model results in Section 4, Table 6]
    "R^f_o := (C^train_online + C^ds_online)/(C^train + C^ds), where C^train_online is the number of online comments of the user in the training set, C^ds_online is the number of online comments of the user in the part of the dataset lost to downsampling. ... Note that C^ds and C^ds_online do not use any data points from the test or validation sets to prevent any form of data leakage."

    The input feature R^f_o is defined directly from the target variable: it counts the user's comments whose Status is 'online', the same binary label the model is trained to predict. Therefore the contextual input is a function of the output it is supposed to predict. All top-scoring deep models in Table 6 include R^f_o or R^s_o, and the paper states that models using the Online Ratio show the best performance. Because the data split is random rather than temporal, a test comment's R^f_o can include the same user's later comments if they landed in the training or downsampled portions, so the feature can encode future moderation outcomes. The stated exclusion of test/validation data prevents direct target-instance leakage but not this cross-instance label leakage.

full rationale

The paper's central claim that context helps LSTM/CNN models is not fully circular: base_LSTM_title and base_LSTM_title_path improve over base_LSTM using article title/path, which are genuinely independent contextual features. However, the largest gains come from models with the label-derived Online Ratio, so the main headline partially reduces to target encoding. No load-bearing self-citation was found; citations to prior datasets and BERT baselines are external. The comparison to previous literature is not circular, though it is not same-data and would need a same-data BERT baseline to strictly support 'competitive with BERT'. Overall, one concrete construction-level circular step warrants a score of 6.

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

The central results rest on the One Million Posts labels as ground truth and on the legitimacy of a user-history feature computed from labels. No new entities are introduced. The absence of reported hyperparameters and split sizes adds hidden choices that affect reproducibility.

assumptions (3)
  • domain assumption The One Million Posts Corpus 'deleted' labels are an accurate ground truth for content moderation decisions under Der Standard's guidelines.
    The paper uses these labels as the binary target without auditing label quality or moderator disagreement; see Section 3.1.
  • domain assumption The online ratio features R^s_o and R^f_o computed from training and downsampled data do not constitute label leakage into the test set.
    The feature uses the actual online/deleted status of the user's other comments; the paper excludes test and validation from the computation, but the feature remains a label-derived statistic; see Section 3.1.
  • domain assumption Randomly dropping the majority of online comments preserves the feature distribution needed for the ratio features and model training.
    Downsampling changes the base rate and may affect user-ratio features; see Section 3.1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Context-Aware Content Moderation for German Newspaper Comments." pith.science (2026). https://pith.science/paper/V4WMYX4H

@misc{pith2026250520963,
  author       = {Pith},
  title        = {Pith review of: Context-Aware Content Moderation for German Newspaper Comments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V4WMYX4H}},
  note         = {Machine review of arXiv:2505.20963}
}
read the original abstract

The increasing volume of online discussions requires advanced automatic content moderation to maintain responsible discourse. While hate speech detection on social media is well-studied, research on German-language newspaper forums remains limited. Existing studies often neglect platform-specific context, such as user history and article themes. This paper addresses this gap by developing and evaluating binary classification models for automatic content moderation in German newspaper forums, incorporating contextual information. Using LSTM, CNN, and ChatGPT-3.5 Turbo, and leveraging the One Million Posts Corpus from the Austrian newspaper Der Standard, we assess the impact of context-aware models. Results show that CNN and LSTM models benefit from contextual information and perform competitively with state-of-the-art approaches. In contrast, ChatGPT's zero-shot classification does not improve with added context and underperforms.

Figures

Figures reproduced from arXiv: 2505.20963 by the authors.

Figure 1
Figure 1. The different deep learning architectures. Traditional Shallow Methods. Traditional shallow models are utilized as base￾line models, with no additional context information, thus predicting the binary classification task solely with the cleaned text of the user comment. After the data preprocessing explained in the last section, it is crucial to transform the text into a numerical representation. For this a Countvect… view at source ↗
Figure 2
Figure 2. Precision-Recall comparison of the different models. AUROC and F1-Score Analysis. The AUROC values for the models mirror the trends observed in accuracy. The advanced LSTM models demonstrate slightly superior AUROC scores compared to other models. In terms of F1-score, the advanced LSTM model using the simple ratio input achieves the highest score. Among the GPT models, GPT_title_strength achieves a relatively high … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

33 extracted references · 18 canonical work pages

  1. [2]

    https://doi.org/10.5281/zenodo.5242915, https: //doi.org/10.5281/zenodo.5242915, [dataset]

    Assenmacher, D., Niemann, M., Müller, K., Seiler, M.V., Riehle, D.M., Traut- mann, H.: RP-Mod & RP-Crowd: Moderator- and Crowd-Annotated German News Comment Datasets (2021). https://doi.org/10.5281/zenodo.5242915, https: //doi.org/10.5281/zenodo.5242915, [dataset]

    work page doi:10.5281/zenodo.5242915 2021

  2. [31]

    In: Proceedings of the 17th Conference on Natural Language Processing (KONVENS 2021)

    Yadav, A., Milde, B.: forumBERT: Topic adaptation and classification of contex- tualized forum comments in German. In: Proceedings of the 17th Conference on Natural Language Processing (KONVENS 2021). pp. 193–202. Düsseldorf, Ger- many (2021), https://aclanthology.org/2021.konvens-1.17/

    work page 2021

  3. [21]

    AustroTox: A Dataset for Target-Based Austrian German Offensive Language Detection

    Pachinger, P., Goldzycher, J., Planitzer, A.M., Kusa, W., Hanbury, A., Neidhardt, J.: Austrotox: A dataset for target-based austrian german offensive language detec- tion. arXiv preprintarXiv:2406.08080(2024), https://arxiv.org/abs/2406.08080

    work page Pith review arXiv 2024

  4. [12]

    In: Information Integration and Web Intelligence, pp

    Keller, M.E., Auch, M., Döschl, A., Vlk, F., Quernheim, J., Hartmann, M., Mandl, P., Kaul, A., Franz, M., Greguš, M., Kotsis, G., Delir Haghighi, P., Khalil, I.: Hocon34k: A corpus of hate speech in online comments from german newspapers. In: Information Integration and Web Intelligence, pp. 212–226. Springer Nature Switzerland (2025). https://doi.org/10....

    work page doi:10.1007/978-3-031-78090-5_18 2025

  5. [1]

    Information13(6) (2022)

    Alkomah, F., Ma, X.: A literature review of textual hate speech detection methods and datasets. Information13(6) (2022). https://doi.org/10.3390/info13060273

    work page doi:10.3390/info13060273 2022

  6. [3]

    Chiu, K.L., Collins, A., Alexander, R.: Detecting hate speech with gpt-3 (2022), https://arxiv.org/abs/2103.12407

    arXiv 2022

  7. [4]

    In: Proceedings of the 7th Workshop on Challenges and Ap- plications of Automated Extraction of Socio-political Events from Text (2024)

    Dehghan, S., Yanikoglu, B.: Evaluating ChatGPT‘s ability to detect hate speech in Turkish tweets. In: Proceedings of the 7th Workshop on Challenges and Ap- plications of Automated Extraction of Socio-political Events from Text (2024). pp. 54–59. Association for Computational Linguistics, St. Julians, Malta (2024), https://aclanthology.org/2024.case-1.6/ C...

    work page 2024

  8. [5]

    ACM Comput

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

    doi:10.1145/3232676 2018

Show all 33 references
  1. [6]

    Neural Computation 9(8), 1735–1780 (1997)

    Hochreiter, S., Schmidhuber, J.: Long short-term memory. Neural Computation 9(8), 1735–1780 (1997). https://doi.org/10.1162/neco.1997.9.8.1735

    1997 doi

  2. [7]

    In: Proceedings of the ACM Web Conference

    Horta Ribeiro, M., Cheng, J., West, R.: Automated content moderation increases adherence to community guidelines. In: Proceedings of the ACM Web Conference

  3. [8]

    In: Proceedings of the 31st International Conference on Computational Linguistics

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

    2025

  4. [9]

    International Journal on Emerging Research Areas (IJERA)04(02), 47–50 (2025)

    Jose,M.,Anthony,J.,Joseph,J.V.,Thomas,J.,Thomas,S.B.:Areviewofmachine learning and deep learning approaches for offensive text detection. International Journal on Emerging Research Areas (IJERA)04(02), 47–50 (2025). https://doi. org/10.5281/zenodo.14651005, https://doi.org/10....

    2025 doi

  5. [10]

    In: Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers

    Joulin, A., Grave, E., Bojanowski, P., Mikolov, T.: Bag of tricks for efficient text classification. In: Proceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers. pp. 427–431. Association for Computatio...

    2017

  6. [11]

    Journal of Management Analytics 7(2), 139–172 (2020)

    Kang,Y.,Cai,Z.,Tan,C.W.,Huang,Q.,Liu,H.:Naturallanguageprocessing(nlp) in management research: A literature review. Journal of Management Analytics 7(2), 139–172 (2020). https://doi.org/10.1080/23270012.2020.1756939

    2020

  7. [13]

    In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)

    Krenn, B., Petrak, J., Kubina, M., Burger, C.: Germs-at: A sexism/misogyny dataset of forum comments from an austrian online newspaper. In: Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024)...

    2024

  8. [14]

    ACM Trans

    Li, L., Fan, L., Atreja, S., Hemphill, L.: “hot” chatgpt: The promise of chatgpt in detecting and discriminating hateful, offensive, and toxic comments on social media. ACM Trans. Web18(2) (2024). https://doi.org/10.1145/3643829

    2024 doi

  9. [15]

    https://doi.org/10.48550/arxiv.2502.06207 (2025), arXiv:2502.06207

    Lu, J., Ma, K., Wang, K., Xiao, K., Lee, R.K.W., Xu, B., Yang, L., Lin, H.: Unveiling the capabilities of large language models in detecting offensive language with annotation disagreement. https://doi.org/10.48550/arxiv.2502.06207 (2025), arXiv:2502.06207

    2025 doi arXiv

  10. [16]

    International Journal of Data Science and Analyt- ics pp

    Malik, J.S., Qiao, H., Pang, G., van den Hengel, A.: Deep learning for hate speech detection: a comparative study. International Journal of Data Science and Analyt- ics pp. 1–16 (2024)

    2024

  11. [17]

    Mikolov, T., Chen, K., Corrado, G., Dean, J.: Efficient estimation of word repre- sentations in vector space (2013), https://arxiv.org/abs/1301.3781

    2013 arXiv

  12. [18]

    Mnassri, K., Farahbakhsh, R., Chalehchaleh, R., Rajapaksha, P., Jafari, A.R., Li, G., Crespi, N.: A survey on multi-lingual offensive language detection. PeerJ. Computer Science10, e1934–e1934 (2024). https://doi.org/10.7717/peerj-cs.1934, https://doi.org/10.7717/peerj-cs.1934...

    2024 doi

  13. [19]

    IEEE Access9, 88364–88376 (2021)

    Mullah, N.S., Zainon, W.M.N.W.: Advances in machine learning algorithms for hate speech detection in social media: A review. IEEE Access9, 88364–88376 (2021). https://doi.org/10.1109/ACCESS.2021.3089515

    2021

  14. [20]

    OpenAI: Openai api (2025), https://mistral.ai/en/news/announcing-mistral-7b

    2025

  15. [22]

    In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP)

    Pennington, J., Socher, R., Manning, C.D.: Glove: Global vectors for word repre- sentation. In: Proceedings of the 2014 conference on empirical methods in natural language processing (EMNLP). pp. 1532–1543 (2014)

    2014

  16. [23]

    Packt Publishing Ltd (2022)

    Raschka, S., Liu, Y.H., Mirjalili, V.: Machine Learning with PyTorch and Scikit- Learn: Develop machine learning and deep learning models with Python. Packt Publishing Ltd (2022)

    2022

  17. [24]

    Engineering Applications of Arti- ficial Intelligence133(Part B), 108159 (2024)

    Saumya, S., Kumar, A., Singh, J.P.: Filtering offensive language from multilingual social media contents: A deep learning approach. Engineering Applications of Arti- ficial Intelligence133(Part B), 108159 (2024). https://doi.org/10.1016/j.engappai. 2024.108159, https://doi.org...

    2024

  18. [25]

    Schabus, D., Skowron, M.: Academic-industrial perspective on the development anddeploymentofamoderationsystemforanewspaperwebsite.In:Proceedingsof the eleventh international conference on language resources and evaluation (2018)

    2018

  19. [26]

    In: Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Schabus, D., Skowron, M., Trapp, M.: One million posts: A data set of german on- line discussions. In: Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval. p. 1241–1244. SIGIR ’17, New York, NY, USA (2017). https://do...

    2017

  20. [27]

    Cyberpsychology: Journal of Psychosocial Research on Cyberspace17(4) (2023)

    Spence, R., Bifulco, A., Bradbury, P., Martellozzo, E., DeMarco, J.: The psycho- logical impacts of content moderation on content moderators: A qualitative study. Cyberpsychology: Journal of Psychosocial Research on Cyberspace17(4) (2023)

    2023

  21. [28]

    arXiv preprint arXiv:1706.03762 (2017), https://doi.org/10.48550/arxiv.1706.03762

    Vaswani,A.,Shazeer,N.,Parmar,N.,Uszkoreit,J.,Jones,L.,Gomez,A.N.,Kaiser, Ł., Polosukhin, I.: Attention is all you need. arXiv preprint arXiv:1706.03762 (2017), https://doi.org/10.48550/arxiv.1706.03762

    2017 doi arXiv

  22. [29]

    PLOS ONE15(12), 1–32 (2021)

    Vidgen, B., Derczynski, L.: Directions in abusive language training data, a sys- tematic review: Garbage in, garbage out. PLOS ONE15(12), 1–32 (2021). https: //doi.org/10.1371/journal.pone.0243300

    2021 doi

  23. [30]

    In: Proceedings of the First Workshop on Abusive Language Online

    Waseem, Z., Davidson, T., Warmsley, D., Weber, I.: Understanding abuse: A typol- ogy of abusive language detection subtasks. In: Proceedings of the First Workshop on Abusive Language Online. pp. 78–84. Association for Computational Linguis- tics, Vancouver, BC, Canada (2017). ...

    2017 doi

  24. [32]

    In: Proceedings of the 13th International Workshop on Semantic Evaluation

    Zampieri, M., Malmasi, S., Nakov, P., Rosenthal, S., Farra, N., Kumar, R.: SemEval-2019 task 6: Identifying and categorizing offensive language in social me- dia (OffensEval). In: Proceedings of the 13th International Workshop on Semantic Evaluation. pp. 75–86. Association for...

    2019 doi

  25. [2023]

    2666–2676

    p. 2666–2676. WWW ’23, Association for Computing Machinery, New York, NY, USA (2023). https://doi.org/10.1145/3543507.3583275

    2023

Pith tools

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