UO-FIE: Combining Exact-Label Supervision with Graded Utility for Factivity Inference
The Factivity Inference Evaluation 2026 (FIE2026) classifies Chinese context-hypothesis pairs into nine ordered factivity intervals. Its evaluation metric rewards both exact predictions and proximity to the correct interval, while 64.1% of the 566 training examples belong to a single class. In preliminary experiments,...
Xing-Chen Xiao
· 0 citations