Language Understanding can be seen as a process of translation from input natural language sentences into commands of a certain Semantic Language that drive the actions associated to the meaning of the sentences. Under this point of view, a new approach is introduced to automatically learn the required transducers from training sets of input-output examples. This approach overcomes certain input-output sequentiality problems of previous techniques. Experiments are presented with a subset of the DARPA ATIS corpus that show the capabilities of the new method to learn useful English-semantic mappings for this task.