Tag
This paper analyzes how localizing edits in human corrections can amplify relative selection bias in model training, providing theoretical insights and empirical evidence from machine translation experiments.