mlsanityData to 7 Oct 2026

Detoxifying Text with MaRCo: Controllable Revision with Experts and Anti-Experts

arXiv 2212.10543 ↗First posted 20 Dec 2022cs.CL cs.AIcited by 36

Not on the map. Why?

Abstract (arXiv)

Text detoxification has the potential to mitigate the harms of toxicity by rephrasing text to remove offensive meaning, but subtle toxicity remains challenging to tackle. We introduce MaRCo, a detoxification algorithm that combines controllable generation and text rewriting methods using a Product of Experts with autoencoder language models (LMs). MaRCo uses likelihoods under a non-toxic LM (expert) and a toxic LM (anti-expert) to find candidate words to mask and potentially replace. We evaluate our method on several subtle toxicity and microaggressions datasets, and show that it not only outperforms baselines on automatic metrics, but MaRCo's rewrites are preferred 2.1 × more in human evaluation. Its applicability to instances of subtle toxicity is especially promising, demonstrating a path forward for addressing increasingly elusive online hate.

Key facts. Cited by 36 papers in the map (data through 7 Oct 2026). Builds on 28 of its 48 references that are in the map. First posted on arXiv 20 Dec 2022.

Builds on

28 of 48 references.

References · 28

  1. 2019RoBERTa: A Robustly Optimized BERT Pretraining Approachcited by 13,266
  2. 2019BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehensioncited by 5,874
  3. 2019BERTScore: Evaluating Text Generation with BERTcited by 5,462
  4. 2019Energy and Policy Considerations for Deep Learning in NLPcited by 1,242
  5. 2020RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Modelscited by 1,235
  6. 2019Green AIcited by 620
  7. 2021DExperts: Decoding-Time Controlled Text Generation with Experts and Anti-Expertscited by 434
  8. 2022ToxiGen: A Large-Scale Machine-Generated Dataset for Adversarial and Implicit Hate Speech Detectioncited by 419
  9. 2019Social Bias Frames: Reasoning about Social and Power Implications of Languagecited by 377
  10. 2017Toward Controlled Generation of Textcited by 271
  11. 2021Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detectioncited by 258
  12. 2017Style Transfer from Non-Parallel Text by Cross-Alignmentcited by 246
16 more
  1. 2020Learning from the Worst: Dynamically Generated Datasets to Improve Online Hate Detectioncited by 217
  2. 2018Delete, Retrieve, Generate: A Simple Approach to Sentiment and Style Transfercited by 191
  3. 2020Reformulating Unsupervised Style Transfer as Paraphrase Generationcited by 181
  4. 2021MultiVerS: Improving scientific claim verification with weak supervision and full-document contextcited by 140
  5. 2021Detoxifying Language Models Risks Marginalizing Minority Voicescited by 106
  6. 2017Best-Worst Scaling More Reliable than Rating Scales: A Case Study on Sentiment Intensity Annotationcited by 77
  7. 2020The Radicalization Risks of GPT-3 and Advanced Neural Language Modelscited by 76
  8. 2020PowerTransformer: Unsupervised Controllable Revision for Biased Language Correctioncited by 59
  9. 2017Shakespearizing Modern Language Using Copy-Enriched Sequence-to-Sequence Modelscited by 57
  10. 2018Fighting Offensive Language on Social Media with Unsupervised Text Style Transfercited by 55
  11. 2020Exploring Controllable Text Generation Techniquescited by 55
  12. 2021Text Detoxification using Large Pre-trained Neural Modelscited by 52
  13. 2019"Mask and Infill" : Applying Masked Language Model to Sentiment Transfercited by 48
  14. 2021Civil Rephrases Of Toxic Texts With Self-Supervised Transformerscited by 32
  15. 2021Artificial intelligence in communication impacts language and social relationshipscited by 22
  16. 2020Contextual Text Style Transfercited by 16

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2023: 10; 2024: 16; 2025: 7; 2026: 3.

36 papers.

Sort:

By area:

  1. 2023Bias and Fairness in Large Language Models: A Surveycited by 626
  2. 2024Detoxifying Large Language Models via Knowledge Editingcited by 86
  3. 2023Controlled Text Generation via Language Model Arithmeticcited by 51
  4. 2025FairSteer: Inference Time Debiasing for LLMs with Dynamic Activation Steeringcited by 37
  5. 2023Goodtriever: Adaptive Toxicity Mitigation with Retrieval-augmented Modelscited by 23
  6. 2024Deconstructing The Ethics of Large Language Models from Long-standing Issues to New-emerging Dilemmas: A Surveycited by 23
  7. 2024SafeInfer: Context Adaptive Decoding Time Safety Alignment for Large Language Modelscited by 22
  8. 2024Mitigating Hallucinations in Large Vision-Language Models via Summary-Guided Decodingcited by 21
  9. 2024Walking in Others' Shoes: How Perspective-Taking Guides Large Language Models in Reducing Toxicity and Biascited by 17
  10. 2023Exploring Methods for Cross-lingual Text Style Transfer: The Case of Text Detoxificationcited by 15
  11. 2023STEER: Unified Style Transfer with Expert Reinforcementcited by 14
  12. 2024MultiParaDetox: Extending Text Detoxification with Parallel Data to New Languagescited by 13
24 more
  1. 2023MacLaSa: Multi-Aspect Controllable Text Generation via Efficient Sampling from Compact Latent Spacecited by 12
  2. 2024Distilling Text Style Transfer With Self-Explanation From LLMscited by 12
  3. 2024Multilingual and Explainable Text Detoxification with Parallel Corporacited by 12
  4. 2023Tackling Bias in Pre-trained Language Models: Current Trends and Under-represented Societiescited by 11
  5. 2024Potential and Challenges of Model Editing for Social Debiasingcited by 10
  6. 2024Effective Controllable Bias Mitigation for Classification and Retrieval using Gate Adapterscited by 8
  7. 2023Learn What NOT to Learn: Towards Generative Safety in Chatbotscited by 7
  8. 2023Don't Take This Out of Context! On the Need for Contextual Models and Evaluations for Stylistic Rewritingcited by 6
  9. 2024MisgenderMender: A Community-Informed Approach to Interventions for Misgenderingcited by 6
  10. 2025Adaptive Detoxification: Safeguarding General Capabilities of LLMs through Toxicity-Aware Knowledge Editingcited by 5
  11. 2024Style Transfer with Multi-iteration Preference Optimizationcited by 4
  12. 2023All Should Be Equal in the Eyes of Language Models: Counterfactually Aware Fair Text Generationcited by 3
  13. 2026CausalDetox: Causal Head Selection and Intervention for Language Model Detoxificationcited by 3
  14. 2024Mitigating Text Toxicity with Counterfactual Generationcited by 2
  15. 2025GradEscape: A Gradient-Based Evader Against AI-Generated Text Detectorscited by 2
  16. 2025Evaluating Text Style Transfer: A Nine-Language Benchmark for Text Detoxificationcited by 2
  17. 2024Contrastive Perplexity for Controlled Generation: An Application in Detoxifying Large Language Modelscited by 1
  18. 2024GrounDial: Human-norm Grounded Safe Dialog Response Generationcited by 1
  19. 2025BanglaNirTox: A Large-scale Parallel Corpus for Explainable AI in Bengali Text Detoxificationcited by 1
  20. 2025C3TG: Conflict-aware, Composite, and Collaborative Controlled Text Generationcited by 1
  21. 2026Do Prompts Guarantee Safety? Mitigating Toxicity from LLM Generations through Subspace Interventioncited by 1
  22. 2024Fine-Grained Detoxification via Instance-Level Prefixes for Large Language Modelscited by 0
  23. 2025Detoxify: A framework for abusive text transformation using LLMscited by 0
  24. 2026Detoxification for LLM: From Dataset Itselfcited by 0