Pith. sign in

REVIEW 4 major objections 5 minor 54 references

Don't Erase, Inform! Detecting and Contextualizing Harmful Language in Cultural Heritage Collections

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

Pith's one-line read A pipeline that flags biased terms in cultural heritage metadata reaches 87 percent precision.

desk verdict A credible, well-documented systems paper with a real deployment and useful public resources; the precision claim is plausible but the evaluation set's construction limits how far the 0.87 transfers to production, and the production annotations are unaudited. read the letter →

arxiv 2505.24538 v1 pith:JRNWUO6P submitted 2025-05-30 cs.CL

classification cs.CL
keywords culturalheritagemetadataharmfullanguagedetectioncontentiousterminologymultilingualvocabularySKOSknowledgegraphLLMdisambiguationEuropeanainclusivecollections
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 claims that an AI pipeline combining a community-built vocabulary, lemmatization, exact string matching, named-entity filtering, and LLM-based disambiguation can flag offensive or outdated terms in cultural heritage metadata at 87 percent precision. The tool's stated purpose is not to erase these terms but to surface them with historical context, so curators and platform users understand why a word is problematic and what alternatives exist. The authors argue this is necessary because heritage metadata is well-formed, historically loaded, and far too voluminous for manual review alone. If the claim holds, heritage institutions gain a deployable mechanism already applied to 7.9 million Europeana records and yielding more than 77,000 annotations.

What carries the argument

The load-bearing mechanism is the DE-BIAS vocabulary, a SKOS knowledge graph that links ContentiousTerm entries to ContentiousIssue descriptions, SuggestionNotes, and SuggestedTerms, with ambiguous terms explicitly flagged. Around it sits a pipeline: tokenization, compound-word splitting for German and Dutch, lemmatization with Stanza, exact string matching against the lemmatized vocabulary, a named-entity recognition filter that suppresses matches occurring inside proper names, and an LLM prompt that decides, given the surrounding text and the vocabulary's own context description, whether an ambiguous occurrence is derogatory. The LLM disambiguation converts a raw string match into a contextual judgment, which is what keeps precision high without burying users in false positives.

What would settle it

Take a random sample of records from Europeana's full corpus without preselecting for any vocabulary terms, run the complete NER-plus-LLM pipeline, and have expert annotators judge every flag; if precision on that sample falls substantially below the reported 0.87, the claim that the tool generalizes to production metadata is refuted.

Watch

Extended reading notes

Core claim

The central claim is that contentious terms in cultural heritage metadata can be detected at scale and with production-usable precision by relying on a curated multilingual vocabulary rather than on a generic toxicity classifier. With both the named-entity filter and the LLM disambiguation module enabled, the tool reaches 0.87 precision on a niche-sourced evaluation set covering all vocabulary terms (Table 1). The vocabulary itself — 687 terms in English, German, Dutch, French, and Italian, linked to 530 contentious issues in a SKOS knowledge graph — is co-created with affected communities and encodes not only which terms are problematic but why, and what wording is suggested instead. Roughly 25 percent of terms are ambiguous and trigger the LLM context check, which the paper identifies as the main false-positive filter. The paper also reports deployment in Europeana's ingestion pipeline and MINT, processing more than 7.9 million records.

Load-bearing premise

The 0.87 precision is measured on an evaluation set that was constructed by first using the tool's own lemmatization and string matching to find candidate texts from 5 million Europeana records, then adding manually gathered texts and a few ChatGPT-generated examples, so the set is skewed toward texts that contain vocabulary terms; if the real distribution of heritage metadata differs, the reported precision may not hold in production.

Editorial extensions

If this is right

  • Cultural heritage platforms can surface biased language across millions of legacy records, something manual curation cannot do at that scale.
  • Because ambiguous terms are flagged only after an LLM check, the pipeline keeps false positives low enough for institutions to trust and adopt it.
  • The published vocabulary, ontology, evaluation dataset, and API give other heritage organizations a reusable starting point for their own detection efforts.
  • The measured throughput of about 813 characters per second with LLM inference makes batch processing of entire collections feasible on a single mid-range GPU.
  • Extending the vocabulary to more languages and terms is the stated next step, and the architecture is designed so new entries plug into the same pipeline.

Reading between the lines

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

  • A natural next test the paper does not run is precision on a random sample of production metadata rather than on texts preselected for vocabulary terms; that would tell how the 0.87 transfers to real collections.
  • The measured language gap (Dutch precision is lowest) suggests that scaling the vocabulary to additional languages may require language-specific LLMs or smaller classifiers rather than a single generic model.
  • The knowledge graph design could support a 'replace-and-link' workflow where suggested alternatives are added alongside original terms, improving searchability; the paper discusses this but did not implement it.
  • If the tool's precision holds in production, the 'inform, not erase' pattern could be adopted by libraries and archives beyond Europeana, since the pattern depends only on having a curated vocabulary and a context-checking 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 a multilingual vocabulary of 687 contentious terms and a detection/contextualization tool for cultural heritage (CH) metadata, combining lemmatization, compound-word splitting, named-entity recognition (NER), and LLM-based disambiguation. The authors report an ablation study with precision 0.87 when both NER and the LLM module are enabled (Table 1), per-language precision across seven LLMs (Table 2), and large-scale deployment at Europeana and MINT, claiming over 7.9 million records processed and more than 77,000 annotations. The evaluation uses a human-annotated dataset built via niche-sourcing, with texts selected from Europeana, supplemented by Wikipedia/dictionaries and ChatGPT-generated examples. The paper openly acknowledges that the evaluation is non-random and does not support recall or accuracy claims.

Significance. If the precision figure transfers to production, the tool provides a practical, community-informed mechanism for surfacing biased language while preserving historical context. The publication of the vocabulary as a SKOS knowledge graph, the evaluation benchmark, and the code are valuable reusable resources for the CH and NLP communities. The co-creation process with affected communities and the human-in-the-loop validation are methodological strengths. However, the transferability of the central precision claim to real CH metadata is not yet established, and the reported deployment metrics do not provide an independent check of that transferability.

major comments (4)
  1. [§5.1 and Table 1] The reported precision of 0.87 is computed on a deliberately balanced, non-random dataset whose candidate texts were retrieved by the tool's own lemmatization and string-matching components from Europeana, supplemented with manually gathered texts from Wikipedia/dictionaries and ChatGPT-generated examples. Because the term distribution is balanced rather than reflecting actual CH metadata, and because the set mixes sources of different genres, Table 1 does not estimate the operational precision an integrator will observe on the full distribution of CH records. The authors should validate on a random sample of the live pipeline's output, or at least report precision separately by source (Europeana vs. external vs. synthetic) and with term-frequency weighting based on production data. Without such a check, the 0.87 figure is a benchmark-specific macro-level precision, not a deployability guarantee.
  2. [§5.1] The use of ChatGPT-generated examples, even after expert review, is an unsupported proxy for real CH metadata. The paper should state how many of the ~3,700 texts are synthetic and report precision separately on the synthetic subset versus the real Europeana/manual subset. If precision is markedly higher on synthetic texts, the aggregate 0.87 is misleading for the CH use case. This is a concrete, testable condition that the authors can meet without new data collection.
  3. [§5.1] Precision-only evaluation is insufficient for a detection tool. The authors themselves write that the targeted approach "did not provide meaningful insights into accuracy or recall." Without recall or F1, a system that flags very few true positives can achieve high precision, so the tool's usefulness for surfacing harmful language is not demonstrated. The paper should report recall or an operating-point analysis on the benchmark, or justify why false negatives are acceptable in the CH context, where the stated goal is to make biases visible.
  4. [§6.2] The claim that the tool processed over 7.9 million records, producing more than 77,000 annotations, is not accompanied by any quality audit of those production annotations. This is a missed opportunity to test transferability of the benchmark precision. The authors should sample and annotate a subset of the production detections (even a few hundred) to verify that the precision measured in the curated evaluation holds in the live pipeline. Without such an audit, the deployment numbers do not support the precision claim beyond the benchmark.
minor comments (5)
  1. [§3.1] The sentence "including euphemisms, dysphemisms, and diminutive" should read "diminutives."
  2. [§3.1 and §Ethical Considerations] There is a typo in "theEthical Considerations" (missing space), and the compensation amounts such as "C200-C300" should be typeset with the euro symbol (e.g., €200–€300) for clarity.
  3. [§5.2 and Table 2] The paper attributes language-level precision differences to the percentage of ambiguous terms per language, but it never reports those percentages; please include this statistic in the text or as a table note so the reader can evaluate the explanation.
  4. [§5.1] The niche-sourcing campaigns involved 145 participants and more than 10,000 votes, yet no inter-annotator agreement metric (e.g., Krippendorff's alpha) is reported; providing one would strengthen the reliability of the resulting benchmark.
  5. [Appendix B] The prompt examples in Table 4 are helpful, but the paper does not describe how the "yes"/"no" model outputs were parsed or how invalid outputs were handled; a sentence on this would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the vocabulary, detection pipeline, and human-labeled evaluation are independent; the evaluation-set selection is a representativeness limitation, not a circular derivation.

full rationale

The derivation chain is not circular. The vocabulary (Section 3) is built from external glossaries, scientific literature, 12 co-creation sessions with over 60 community members, and term-occurrence checks in CH platforms; it is not derived from the tool's output. The detection pipeline (Section 4) is a conventional lemmatization plus exact string match against this externally constructed vocabulary, with Stanza NER filtering and a zero-shot LLM disambiguation that receives the vocabulary's contextual description and must decide yes/no on the text. No parameter is fitted to the evaluation set. The evaluation (Section 5.1) does use the tool's own lemmatization and string-matching components to select candidate texts from 5 million Europeana records, plus Wikipedia, dictionary, and ChatGPT supplements, but the gold labels come from 145 niche-sourcing participants casting over 10,000 up/down votes, with manual organizer review for unclear consensus. Thus the precision 0.87 in Table 1 is a human-validated rate on detected candidates, not the tool's own output compared with itself. The non-random, term-balanced construction limits recall claims and external validity—the paper itself states it 'did not provide meaningful insights into accuracy or recall'—but that is a representativeness limitation, not circular reasoning. Self-citations (the CrowdHeritage platform papers and the DE-BIAS technical documentation) describe infrastructure or project resources and are not load-bearing for the detection claim; the LLM disambiguation is zero-shot and externally available, and the benchmark is released publicly. No step reduces, by construction or by self-citation, to its own inputs.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests primarily on the co-created vocabulary as ground truth and on the reliability of Stanza and LLMs rather than on fitted numerical parameters. The hand-chosen elements are the consensus threshold, custom lemmatization rules, and quantization choices. No new physical or conceptual entities are introduced beyond the vocabulary and KG structure, which are data resources rather than theoretical constructs.

free parameters (3)
  • Consensus threshold for manual review = absolute vote difference < 2
    In the niche-sourcing annotation, cases where the absolute difference between upvotes and downvotes is less than two were manually reviewed; this threshold affects which labels are human-adjudicated and thus the final dataset.
  • Custom lemmatization rules = 162 rules
    Hand-crafted Stanza rules to correct lemmatization for rare and antiquated vocabulary terms; these rules are choices made by the authors and can affect detection.
  • Mixtral quantization bits for ablation = 3 bits
    The ablation study used Mixtral-8x7B quantized to 3 bits; quantization changes model outputs and therefore precision measurements.
assumptions (5)
  • domain assumption The vocabulary terms and their contentiousness labels are a valid ground truth for harmful language in CH metadata.
    The entire evaluation treats the co-created vocabulary as the reference; the paper acknowledges that validation did not cover all languages and themes (Limitations section).
  • domain assumption Stanza lemmatization with the 162 custom rules produces correct canonical forms for all supported languages.
    The string-matching detection relies on lemmatized forms; incorrect lemmas would cause misses (low recall) or false matches.
  • domain assumption The LLM's yes/no answer about contentiousness is a reliable indicator when prompted with the vocabulary context.
    The LLM disambiguation module is a zero-shot component; its reliability is only indirectly probed through the aggregate precision numbers.
  • ad hoc to paper ChatGPT-generated example texts, after expert review, are representative of real CH metadata.
    In Section 5.1, ChatGPT was used to generate supplementary examples for terms with few occurrences; these are synthetic and may differ in style from actual CH records.
  • domain assumption The community co-creation process yields a culturally appropriate and consistent set of contentious terms across five languages.
    The vocabulary is the product of 12 co-creation sessions with over 60 participants; the paper states that full validation across all themes and languages was not feasible.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Don't Erase, Inform! Detecting and Contextualizing Harmful Language in Cultural Heritage Collections." pith.science (2026). https://pith.science/paper/JRNWUO6P

@misc{pith2026250524538,
  author       = {Pith},
  title        = {Pith review of: Don't Erase, Inform! Detecting and Contextualizing Harmful Language in Cultural Heritage Collections},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JRNWUO6P}},
  note         = {Machine review of arXiv:2505.24538}
}
read the original abstract

Cultural Heritage (CH) data hold invaluable knowledge, reflecting the history, traditions, and identities of societies, and shaping our understanding of the past and present. However, many CH collections contain outdated or offensive descriptions that reflect historical biases. CH Institutions (CHIs) face significant challenges in curating these data due to the vast scale and complexity of the task. To address this, we develop an AI-powered tool that detects offensive terms in CH metadata and provides contextual insights into their historical background and contemporary perception. We leverage a multilingual vocabulary co-created with marginalized communities, researchers, and CH professionals, along with traditional NLP techniques and Large Language Models (LLMs). Available as a standalone web app and integrated with major CH platforms, the tool has processed over 7.9 million records, contextualizing the contentious terms detected in their metadata. Rather than erasing these terms, our approach seeks to inform, making biases visible and providing actionable insights for creating more inclusive and accessible CH collections.

Figures

Figures reproduced from arXiv: 2505.24538 by the authors.

Figure 1
Figure 1. A simplified version of the KG structure [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Tool Architecture Overview only matched with text spans that contain the ex￾act same tokens (after lemmatization) in the exact same order. 4.3 NER To prevent false detections of vocabulary terms within named entities, we integrated a Named En￾tity Recognition (NER) module into our pipeline. One common case is terms appearing as surnames, e.g. “Sordo” in Italian. Using the NER module from Stanza, known for its robust… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 35 canonical work pages

  1. [1]

    Pedro Alonso, Rajkumar Saini, and Gy \"o rgy Kov \'a cs. 2020. Hate speech detection using transformer ensembles on the hasoc dataset. In International conference on speech and computer, pages 13--21. Springer

  2. [2]

    Thomas Baker, Sean Bechhofer, Antoine Isaac, Alistair Miles, Guus Schreiber, and Ed Summers. 2013. Key choices in the design of simple knowledge organization system (skos). Journal of Web Semantics, 20:35--49

  3. [3]

    Michele Banko, Brendon MacKeen, and Laurie Ray. 2020. https://doi.org/10.18653/v1/2020.alw-1.16 A unified taxonomy of harmful content . In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 125--137, Online. Association for Computational Linguistics

  4. [4]

    Marco Bellagente, Jonathan Tow, Dakota Mahan, Duy Phung, Maksym Zhuravinskyi, Reshinth Adithyan, James Baicoianu, Ben Brooks, Nathan Cooper, Ashish Datta, et al. 2024. Stable lm 2 1.6 b technical report. arXiv preprint arXiv:2402.17834

  5. [5]

    Michele Bevilacqua, Tommaso Pasini, Alessandro Raganato, and Roberto Navigli. 2021. Recent trends in word sense disambiguation: A survey. In International Joint Conference on Artificial Intelligence, pages 4330--4338. International Joint Conference on Artificial Intelligence, Inc

  6. [6]

    Meghana Moorthy Bhat, Saghar Hosseini, Ahmed Hassan, Paul Bennett, and Weisheng Li. 2021. Say ‘yes’to positivity: Detecting toxic language in workplace communications. In Findings of the Association for Computational Linguistics: EMNLP 2021, pages 2017--2029

  7. [7]

    Tom Bourgeade, Zongmin Li, Farah Benamara, V \'e ronique Moriceau, Jian Su, and Aixin Sun. 2024. https://aclanthology.org/2024.lrec-main.740 Humans need context, what about machines? investigating conversational context in abusive language detection . In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resource...

  8. [8]

    Peter Burnap and Matthew Leighton Williams. 2014. Hate speech, machine classification and statistical modelling of information flows on twitter: Interpretation and communication for policy decision making

Show all 54 references
  1. [9]

    Tuhin Chakrabarty, Kilol Gupta, and Smaranda Muresan. 2019. https://doi.org/10.18653/v1/W19-3508 Pay `` attention '' to your context when classifying abusive language . In Proceedings of the Third Workshop on Abusive Language Online, pages 70--79, Florence, Italy. Association ...

  2. [10]

    Kimberl \'e Williams Crenshaw. 2013. Mapping the margins: Intersectionality, identity politics, and violence against women of color. In The public nature of private violence, pages 93--118. Routledge

  3. [11]

    Sam Davidson, Qiusi Sun, and Magdalena Wojcieszak. 2020. https://doi.org/10.18653/v1/2020.alw-1.12 Developing a new classifier for automated identification of incivility in social media . In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 95--101, Online. A...

  4. [12]

    Thomas Davidson, Debasmita Bhattacharya, and Ingmar Weber. 2019. https://doi.org/10.18653/v1/W19-3504 Racial bias in hate speech and abusive language detection datasets . In Proceedings of the Third Workshop on Abusive Language Online, pages 25--35, Florence, Italy. Associatio...

  5. [13]

    DE-BIAS Project . 2024. The DE-BIAS vocabulary and knowledge graph - documentation. https://pro.europeana.eu/files/Europeana_Professional/Projects/de_bias_vocabulary/TechnicalDocumentation_DE-BIAS_VocabularyKnowledgeGraph.pdf. Accessed: 2025-05-30

  6. [14]

    Fabio Del Vigna, Andrea Cimino, Felice Dell’Orletta, Marinella Petrocchi, and Maurizio Tesconi. 2017. Hate me, hate me not: Hate speech detection on facebook. In Proceedings of the first Italian conference on cybersecurity (ITASEC17), pages 86--95

  7. [15]

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

  8. [16]

    AbdelRahim Elmadany, Chiyu Zhang, Muhammad Abdul-Mageed, and Azadeh Hashemi. 2020. https://aclanthology.org/2020.osact-1.17 Leveraging affective bidirectional transformers for offensive language detection . In Proceedings of the 4th Workshop on Open-Source Arabic Corpora and P...

  9. [17]

    European Commission , Directorate-General for Justice, Consumers, and L Kaati. 2023. https://doi.org/doi/10.2838/59442 The European online hate lab . Publications Office of the European Union

  10. [18]

    Albert Q Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, et al. 2024. Mixtral of experts. arXiv preprint arXiv:2401.04088

  11. [19]

    Eirini Kaldeli, Orfeas Menis-Mastromichalakis, Spyros Bekiaris, Maria Ralli, Vassilis Tzouvaras, and Giorgos Stamou. 2021. Crowdheritage: crowdsourcing for improving the quality of cultural heritage metadata. Information, 12(02):64

  12. [20]

    Divyansh Kaushik, Eduard Hovy, and Zachary Lipton. 2020. https://openreview.net/forum?id=Sklgs0NFvr Learning the difference that makes a difference with counterfactually-augmented data . In International Conference on Learning Representations

  13. [21]

    Brendan Kennedy, Xisen Jin, Aida Mostafazadeh Davani, Morteza Dehghani, and Xiang Ren. 2020. https://doi.org/10.18653/v1/2020.acl-main.483 Contextualizing hate speech classifiers with post-hoc explanation . In Proceedings of the 58th Annual Meeting of the Association for Compu...

  14. [22]

    Katerina Korre, John Pavlopoulos, Jeffrey Sorensen, L \'e o Laugier, Ion Androutsopoulos, Lucas Dixon, and Alberto Barr \'o n-Cede \ n o. 2023. Harmful language datasets: An assessment of robustness. In The 7th Workshop on Online Abuse and Harms (WOAH), pages 221--230

  15. [23]

    Jana Kurrek, Haji Mohammad Saleem, and Derek Ruths. 2020. https://doi.org/10.18653/v1/2020.alw-1.17 Towards a comprehensive taxonomy and large-scale annotated corpus for online slur usage . In Proceedings of the Fourth Workshop on Online Abuse and Harms, pages 138--149, Online...

  16. [24]

    Daniel Loureiro, Kiamehr Rezaee, Mohammad Taher Pilehvar, and Jose Camacho-Collados. 2021. Analysis and evaluation of language models for word sense disambiguation. Computational Linguistics, 47(2):387--443

  17. [25]

    Pedro Henrique Martins, Patrick Fernandes, Jo \ a o Alves, Nuno M Guerreiro, Ricardo Rei, Duarte M Alves, Jos \'e Pombal, Amin Farajian, Manuel Faysse, Mateusz Klimaszewski, et al. 2024. Eurollm: Multilingual language models for europe. arXiv preprint arXiv:2409.16235

  18. [26]

    Anneleen Masschelein, Frederik Truyen, Sofie Taes, Jonas Van Mulder, Andreas Stynen, and Roberta Pireddu. 2024. https://doi.org/https://doi.org/10.5281/zenodo.14514378 Bias types and patterns: a typology applied to europeana use cases

  19. [27]

    Puneet Mathur, Rajiv Shah, Ramit Sawhney, and Debanjan Mahata. 2018. https://doi.org/10.18653/v1/W18-3504 Detecting offensive tweets in H indi- E nglish code-switched language . In Proceedings of the Sixth International Workshop on Natural Language Processing for Social Media,...

  20. [28]

    o rn Gamb \

    Johannes Skjeggestad Meyer and Bj \"o rn Gamb \"a ck. 2019. https://doi.org/10.18653/v1/W19-3516 A platform agnostic dual-strand hate speech detector . In Proceedings of the Third Workshop on Abusive Language Online, pages 146--156, Florence, Italy. Association for Computation...

  21. [29]

    Wayne Modest. 2018. Words Matter: an unfinished guide to word choices in the cultural sector. Tropenmuseum, Afrikamuseum, Museum Volkenkunde, Wereldmuseum

  22. [30]

    Sandip Modha, Prasenjit Majumder, and Thomas Mandl. 2018. https://aclanthology.org/W18-4423 Filtering aggression from the multilingual social media feed . In Proceedings of the First Workshop on Trolling, Aggression and Cyberbullying ( TRAC -2018) , pages 199--207, Santa Fe, N...

  23. [31]

    Andrei Nesterov, Laura Hollink, Marieke van Erp, and Jacco van Ossenbruggen. 2023. A knowledge graph of contentious terminology for inclusive representation of cultural heritage. In European Semantic Web Conference, pages 502--519. Springer

  24. [32]

    Niraula, Saurab Dulal, and Diwa Koirala

    Nobal B. Niraula, Saurab Dulal, and Diwa Koirala. 2021. https://doi.org/10.18653/v1/2021.woah-1.7 Offensive language detection in N epali social media . In Proceedings of the 5th Workshop on Online Abuse and Harms (WOAH 2021), pages 67--75, Online. Association for Computationa...

  25. [33]

    Flor Miriam Plaza-del Arco, Debora Nozza, Dirk Hovy, et al. 2023. Respectful or toxic? using zero-shot learning with language models to detect hate speech. In The 7th Workshop on Online Abuse and Harms (WOAH). Association for Computational Linguistics

  26. [34]

    Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and Christopher D. Manning. 2020. https://nlp.stanford.edu/pubs/qi2020stanza.pdf Stanza: A Python natural language processing toolkit for many human languages . In Proceedings of the 58th Annual Meeting of the Association for Co...

  27. [35]

    Maria Ralli, Spyros Bekiaris, Eirini Kaldeli, Orfeas Menis-Mastromichalakis, Natasa Sofou, Vassilis Tzouvaras, and Giorgos Stamou. 2020. Crowdheritage: Improving the quality of cultural heritage through crowdsourcing methods. In 2020 15th International Workshop on Semantic and...

  28. [36]

    Haji Mohammad Saleem, Jana Kurrek, and Derek Ruths. 2022. Enriching abusive language detection with community context. In Proceedings of the Sixth Workshop on Online Abuse and Harms (WOAH), pages 131--142

  29. [37]

    Maarten Sap, Dallas Card, Saadia Gabriel, Yejin Choi, and Noah A. Smith. 2019. https://doi.org/10.18653/v1/P19-1163 The risk of racial bias in hate speech detection . In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1668--1678, ...

  30. [38]

    A. Sasse. 2015. https://doi.org/10.1109/MSP.2015.65 Scaring and bullying people into security won't work . IEEE Security & Privacy, 13:80--83

  31. [39]

    Indira Sen, Mattia Samory, Claudia Wagner, and Isabelle Augenstein. 2022. https://doi.org/10.18653/v1/2022.naacl-main.347 Counterfactually augmented data and unintended bias: The case of sexism and hate speech detection . In Proceedings of the 2022 Conference of the North Amer...

  32. [40]

    Hyeonho Song, Jisu Hong, Chani Jung, Hyojin Chin, Mingi Shin, Yubin Choi, Junghoi Choi, and Meeyoung Cha. 2024. Detecting offensive language in an open chatbot platform. In Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources ...

  33. [41]

    Melanie Subbiah, Amrita Bhattacharjee, Yilun Hua, Tharindu Kumarage, Huan Liu, and Kathleen McKeown. 2023. https://doi.org/10.18653/v1/2023.wassa-1.11 Towards detecting harmful agendas in news articles . In Proceedings of the 13th Workshop on Computational Approaches to Subjec...

  34. [42]

    TGDK Sumanathilaka, Nicholas Micallef, and Julian Hough. 2024. Can llms assist with ambiguity? a quantitative evaluation of various large language models on word sense disambiguation. arXiv preprint arXiv:2411.18337

  35. [43]

    Dias Oliva Thiago, Antonialli Dennys Marcelo, and Alessandra Gomes. 2021. Fighting hate speech, silencing drag queens? artificial intelligence in content moderation and risks to lgbtq voices online. Sexuality & culture, 25(2):700--732

  36. [44]

    St \'e phan Tulkens, Lisa Hilte, Elise Lodewyckx, Ben Verhoeven, and Walter Daelemans. 2016. A dictionary-based approach to racism detection in dutch social media. arXiv preprint arXiv:1608.08738

  37. [45]

    William Warner and Julia Hirschberg. 2012. https://aclanthology.org/W12-2103 Detecting hate speech on the world wide web . In Proceedings of the Second Workshop on Language in Social Media, pages 19--26, Montr \'e al, Canada. Association for Computational Linguistics

  38. [46]

    Zeerak Waseem and Dirk Hovy. 2016. https://doi.org/10.18653/v1/N16-2013 Hateful symbols or hateful people? predictive features for hate speech detection on T witter . In Proceedings of the NAACL Student Research Workshop , pages 88--93, San Diego, California. Association for C...

  39. [47]

    Michael Wiegand, Josef Ruppenhofer, and Thomas Kleinbauer. 2019. Detection of abusive language: the problem of biased datasets. In Proceedings of the 2019 conference of the North American Chapter of the Association for Computational Linguistics: human language technologies, vo...

  40. [48]

    Mengzhou Xia, Anjalie Field, and Yulia Tsvetkov. 2020. https://doi.org/10.18653/v1/2020.socialnlp-1.2 Demoting racial bias in hate speech detection . In Proceedings of the Eighth International Workshop on Natural Language Processing for Social Media, pages 7--14, Online. Assoc...

  41. [49]

    Jung H Yae, Nolan C Skelly, Neil C Ranly, and Phillip M LaCasse. 2024. Leveraging large language models for word sense disambiguation. Neural Computing and Applications, pages 1--18

  42. [50]

    Tsungcheng Yao, Ernest Foo, and Sebastian Binnewies. 2024. Personalised abusive language detection using llms and retrieval-augmented generation. In Proceedings of the 7th International Conference on Natural Language and Speech Processing (ICNLSP 2024), pages 92--98

  43. [51]

    Guanhua Zhang, Bing Bai, Junqi Zhang, Kun Bai, Conghui Zhu, and Tiejun Zhao. 2020. https://doi.org/10.18653/v1/2020.acl-main.380 Demographics should not be the reason of toxicity: Mitigating discrimination in text classifications with instance weighting . In Proceedings of the...

  44. [52]

    Li Zhou, Antonia Karamolegkou, Wenyu Chen, and Daniel Hershcovich. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.845 Cultural compass: Predicting transfer learning success in offensive language detection with cultural features . In Findings of the Association for Compu...

  45. [53]

    URL: " 'urlintro :=

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

  46. [54]

    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 gl...

Pith tools

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