From ML to MLF: graphic type constraints with efficient type inference