The Maskability Index: Predicting Task-Objective Alignment in Pretrained Language Models
A new metric estimates when a relation fits masked prompting better than prefix prompting.
Pouramini and Afsharzadeh define the Maskability Index from DepthRank differences between masked and unmasked templates. They test it on relations from the ATOMIC2020 knowledge-base completion benchmark. The paper reports a positive correlation between MI and downstream generation performance, positioning it as a guide for prompt-template and adaptation choices in low-resource relation extraction. ArXiv · AI/CL/LG's note
Pouramini and Afsharzadeh define the Maskability Index from DepthRank differences between masked and unmasked templates. They test it on relations from the ATOMIC2020 knowledge-base completion benchmark. The paper reports a positive correlation between MI and downstream generation performance, positioning it as a guide for prompt-template and adaptation choices in low-resource relation extraction. ArXiv · AI/CL/LG's note
score 4