He's not assigning entropy per word. He's using NIST SP 800-63, based on the character-length of the phrase, to assign the entropy value.

Proof: https://news.ycombinator.com/item?id=7163362