To get 128 bits of entropy with words, you need to pick about thirteen out of a million words--which is on the order of all the words in the English language--and give all of them equal probability. The sequence needs to be fully random as well. What you end up with will surely be easier to memorize than a UUID, but substantially more difficult than the start of the Lord's Prayer.

EDIT: Math is wrong, I was thinking 10 bits per million instead of 20. So 6-7 words out of a million (whole language) or 13 words out of a thousand (very limited subset of the language). Point about random selection still stands, but it's certainly easier than 13 very uncommon words. Still much harder than a realistic sentence of that length, though.

I'm sure that even with a great statistical model of password haikus (say an LLM) yours would still be one in a billion which still seems unlikely, but a cracking cluster can try billions per second.

In these cases it's very easy to have security that depends on the odds that a powerful attacker just hasn't gotten around to seriously trying the broad class of predictable generation schemes you've used.