Robustness of LLMs to Perturbations in Text