Towards Robust and Accurate Stability Estimation of Local Surrogate Models in Text-based Explainable AI