Macro forest transducers

XML documents conceptually are not trees, but forests. Therefore, we extend the concept of macro tree transducers (mtt's) to a transformation formalism of forests, macro forest transducers (mft's). We show that mft's form a strict superset of mtt's operating on forests (represented as binary trees). On the other hand, the transformation of every mft can be simulated by the composition of two mtt's. Although macro forest transducers are more powerful, the complexity of inverse type inference, i.e., computing the pre-image of a recognizable language, is almost the same as for tree transducers.