Abstract: Textual adversarial samples are widely used to assess the robustness and security of language models. Most existing methods generate these samples by substitution or deletion. However, such ...