I think the problem you are having is caused in preprocessing of prompts. If your prompt is written in a supported language other than English, the entire prompt is translated to English before being given to the model for generation. This ends up translating any text you wanted to have added in the image as well.
There is no easy, clean way around this. You can use a composition reference (in Firefly). You can try writing your prompt in English except for the words you want to have displayed, which might prevent the full prompt translation from occurring. You also can try adding punctuation in the middle of words so the translation process does not recognise the word, and then later remove the punctuation with FireflyGenerative Fillor Photoshop. You also can use Adobe Express or Photoshop to add the text after generation.