It's a big strength of embeddings like BERT and later on that they embed not just the "word" but the word plus its context so if the same word is used in two different places with two different meanings the embedding for that token ought to be different.