How is the similarity between synonyms or semantically similar sentences ensured if regex is used for retrieving the input prompts? Maybe I missed something as I skimmed over the paper, but that was the impression I got