[lhnz]

I've heard that nowadays subword/token embeddings are learned during the training phase, and that they are useful for reconstructing the embeddings of words that contain them, and in fact allow the model to handle typos like "aple" (instead of "apple").