The Maskability Index: Predicting Task-Objective Alignment in Pretrained Language Models

This paper introduces the Maskability Index (MI), a metric for estimating whether a relational knowledge task is better matched to masked-style or prefix-style prompting. MI is computed from differences in DepthRank scores between masked and unmasked templates. The authors evaluate it across relations in the ATOMIC2020 knowledge-base completion benchmark and report a positive correlation between MI and downstream generation performance. The proposed metric is intended to guide prompt-template selection and adaptation strategies, particularly in few-shot and low-resource settings involving pretrained models such as T5 and BERT.