The generator must be given some preferences on the above options in order to help it make decisions. For instance, the generator will prefer to always use grammatical constraints as in (8), where to keep the same meaning as in (8-a), one should select (8-c) with a COMP and not (8-b) with the XCOMP. The generator will also prefer collocations, as said previously. Some other preferences, still lexicon based, will be provided by the stylistic and frequency features of an entry, such as preferring (10-a) for a formal text over (10-b) which is less frequent or over (10-c) which is not formal.
Finally, the generator must allow for dealing with combination-based preferences or in other words, looking at interactions globally and not only locally. For instance, the global stylistics of a text might be different from the local stylistic of lexemes looked at individually. So, assuming we want a formal text, the generator must keep track of how many very formal lexemes we have used, as we would not want to end up with a strange text. Some other global preferences will be goal driven, such as maximizing the coreference possibilities.