What:

This is literally just summing up the one-hot encodings into a single vector. That’s it. (Obviously doesn’t care about the order of words - but still surprisingly works for topic classification.)