I wonder 'where' these compound words end up in an n-dim embedding space (relative to their German and say English 'parts'). In fact this brings up the interesting question of tokenization of the long German compound words, and how all this plays out in German to English (and reverse) LLM translation and text generation.