Actually the generation of the word lists from a given entropy does not increase it, if the dictionary is known and fixed,  just like hashing does not.
Therefore the key set is size is determined by the entropy generator. I was overestimating the entropy using the stats of the language.