As an example of a rule of such a transfer grammar, consider the translation of the following pair of sentences of English and Dutch:
It is assumed that the Dutch semantic structure takes John to be the first, and Mary the second argument. In the English semantic structure the situation is reversed. The following transfer rule might be written to translate such structures into each other. The labels nl and gb refer to the Dutch resp. the English semantic structure.
It is discussed how a simple constraint on such transfer grammars can be defined, to guarantee that transfer is effective. This constraint requires that for a given transfer rule between languages l1 and l2, the value of the l1 attribute of the mother node is strictly larger than each of the values for this attribute of the daughters. The same condition holds for the l2 attribute. It can easily be shown that for a grammar whose rules adhere to this condition, termination is guaranteed for all inputs.
It might be expected that such a constraint reduces the expressive power of a transfer grammar. It is shown, by means of an example, that powerful feature percolations can be used to implement certain `non-compositional' translation cases. It can thus be argued that reversible transfer grammars in fact constitute an interesting compromise between expressive power and computability.