New AEGIS framework explores span-guided multilingual text detoxification

A new framework called AEGIS has been developed to study the effectiveness of span-level guidance in multilingual text detoxification. This framework combines span detector outputs with frozen generator backbones, allowing for the provision of harmful spans, intensity labels, and target attributes as structured guidance during the rewriting process. The research indicates that while span-guided detoxification can alter the trade-off between reducing toxicity and preserving meaning, its impact is highly dependent on the specific generator backbone, model scale, and language used. AI

IMPACT Provides insights into controlling AI-generated text and improving safety across multiple languages.

RANK_REASON The item is an academic paper detailing a new framework for studying text detoxification. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• AEGIS
• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• English
• Gotit.pub
• Hugging Face
• Korean
• ScienceCast
• Standard Chinese

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →