Natural language generation is translated to a planning problem in order to benefit from off-the-shelf planners. The efficiency
of FF and SGPLAN is tested with the sentence generation problem as well as the instruction giving domain and
compared to other conventional tools. Experiments have been made using different configurations. An important amount
of time of the total running time has been allocated to the grounding/parsing step which is not necessarily useful in this
application. To solve this problem for the FF planner, some fixes has been performed. As a first step to make the preprocessing
phase less exhausting, hard processors have been eliminated. Second, action pruning and helpful actions techniques
have been adopted.