Apple ML Research: Annotation needs vary by evaluation metric

Apple Machine Learning Research has published a paper detailing a method called Metric-Dependent Annotation Saturation. This approach suggests that the number of annotators required to capture meaningful signal from label distributions is dependent on the specific evaluation metric being used. For instance, achieving convergence for entropy correlation in NLI models requires significantly more annotators than for distributional match. The research also highlights that soft labels, which represent nuanced decision boundaries, offer better regularization and generalization than one-hot labels, especially when dealing with noisy annotations. AI

IMPACT Suggests optimizing annotation budgets based on evaluation metrics for improved model training.

RANK_REASON Research paper published by Apple's ML Research division. [lever_c_demoted from research: ic=1 ai=1.0]

Read on Apple Machine Learning Research →

• Apple Machine Learning Research
• ChaosNLI
• DeBERTa
• Guneet Kohli
• Metric-Dependent Annotation Saturation for Learning from Label Distributions
• RoBERTa

• paper
• model release

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