But does it actually add entropy when a hacker could use a dictionary and combine those words in various ways? The best practice of using several random words is still 'rule based' - the individual 'units' in the password simply become words instead of characters, and the arbitrary length doesn't really matter. Start with the most common 40,000 words in English, and combine them in all possible orders - that gives 2.6e18 combinations. Compare that to the "random" password of length 10, with say, 40 possible characters: 1e16 combinations. I think OP has a point about the relative strength.
Add a few fun prefixes and suffixes hither and yon, and you largely eliminate the "token" nature of words as well. Even a couple of well-placed (but ordinarily inappropriate) uns, antis, disens, ousitys and ishnesses increase the problem space dramatically without significantly decreasing (and perhaps even increasing) the memorability.
