Reproduce “Mitigating Text Toxicity with Counterfactual Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Milan Bhan, Jean-Noel Vittaut, Nina Achache, Victor Legrand, Nicolas Chesneau, Annabelle Blangero, Juliette Murris, Marie-Jeanne Lesot
Researchers from Ekimetrics, Sorbonne Université, Aix-Marseille Université, and Université Paris Cité developed CF-Detox_tigtec_, a text detoxification method that leverages eXplainable AI (XAI) techniques. The method demonstrated superior content preservation and competitive toxicity reduction across multiple datasets, balancing these metrics effectively against existing approaches.
Key insights
- Counterfactual generation, typically an XAI technique for explanation, can be effectively adapted as an active method for text transformation tasks such as detoxification.
- The proposed "target-then-replace-then-refine" framework offers a structured, model-driven approach that addresses the critical challenge of content preservation in text detoxification.
- The CF-Detox_tigtec_ framework is robust to the choice of different LFI methods (e.g., KernelSHAP, Integrated gradients, Self-attention), indicating flexibility in implementation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.