With LLM, the dictionaries come from the training data, and training also says what's the most probable link between a word and the next word.