Researchers have identified a flaw in existing formality transfer datasets, such as GYAFC, where human rewrites encode relative stylistic shifts rather than absolute formality. This leads models to generate outputs that satisfy benchmark labels but are not genuinely formal. To address this, a new framework is proposed that views formality as a graded dimension with three levels: informal, casual, and formal, with 'casual' acting as an intermediate state. A new dataset, 3LF, has been created based on this framework, which significantly improves model performance in informal-to-formal transfer and better aligns with human perception. AI
IMPACT Introduces a new dataset and framework that improves model alignment with human perception in formality transfer tasks.
RANK_REASON The cluster contains an academic paper introducing a new dataset and framework for a specific NLP task. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
