Without knowing the specific details of the mathematics of entropy, I kinda get the feeling that grammatically correct sentences are not a good source of entropy. Consider that sentence structure in a given language is, deliberately, an inherent restriction in the way that different words can be combined in a sequence.
The nice thing about mnemonic techniques is that you know, with certainty, that they contain enough entropy because you can put a big number in, and get the same number back out again and you're secure assuming that the number was chosen uniformly over a large enough range to satisfy standard cryptographic assumptions. Given that, you're free to optimize for memorability.
It wouldn't be terribly hard to impose a structural constraint— e.g. split your dictionary into parts of speech and have several permitted sentence schemas pick which dictionaries are used in which position, but it might add a couple words to the length.
Another possibility would be to omit verbs and connective words (a, an, the, etc) from the dictionary and let you type them freely but they're ignored by the algorithm.
