Subword models struggle with word learning, but surprisal hides it

Open in new window