Reliability-Aware Sexism Detection: Combining DPO with Annotator Agreement and Token-Level Confidence Scoring
The episode discusses a paper titled "Reliability-Aware Sexism Detection," which combines DPO with annotator agreement and token-level confidence scoring. The authors propose a unified framework for AI that is reliable in subjective tasks like sexism detection. Key findings include efficient training using only the most reliable data and implementing principled abstention at inference, resulting in significantly higher accuracy.