Conditional Partner Selection Recovers 82% of Optimal Span Tagger Gains

Selecting a second model based on held-out accuracy on a primary model's specific errors recovers 82% of the best partner's gain. Using the second model's predictive distribution rather than hard labels improves 94% of pairings for selective span tagging.