When the linguistic knowledge of an NLS increases (especially in the case of lexical and semantic information) a system using reversible knowledge sources is less redundant and hence more efficient than a system that has to manage different sources for understanding and generation. In order to avoid inconsistencies of the language behaviour of the system as mentioned above it is necessary to update all different sources before the system can use it. The disadvantage of using different grammars and lexicons is not only that the knowledge has to be specified twice, but also that the consistency of the new knowledge with respect to the old one has to be checked in many different places. Using a reversible grammar these problems are easily avoided. For example if a user uses a word which is unknown to the system, then acquiring this word during the understanding mode of the system means that this word is automatically available for generation. Clearly, knowledge acquisition in this case presupposes that the knowledge is represented declaratively.