The basic motivation is that vocabularies are unbounded, but embedding tables can only be a fixed size. If you know your training data matches up to your test data well, this isn't such a big deal for most applications. If it's rare at training time, it'll be rare at test time.

→ View original post on X — @honnibal