Assessing reliability of BERT-based models on question answering tasks
This study evaluates the reliability of four BERT-based models by assessing response stability under two conditions: internal model variations induced via Monte Carlo Dropout (MCD) and input perturbations through paraphrasing.