We have presented a new architecture capable of near linear-time, context dependent, multilingual planning. We have demonstrated the advantages of such a system using the case of semantic mismatch. Semantic mismatch presents considerable problems to generation systems that do not rely primarily on semantics. We have shown how correct meanings can be generated only by considering the semantic interactions of various words in a sentence. The PICARD system is able to efficiently plan such sentences by connecting localized means-ends plans with usage constraints and processing the global solution using efficient constraint-based techniques. We have also demonstrated that the same techniques allow for an unlimited amount of interaction between the intelligent planning and linguistic realization components. Of course, the utility of such interaction will only be fully realized when the phenomena is studied more closely and a language of interaction is developed. We hope that the system presented here is a step in that direction, in that the ability to process such systems using large-scale knowledge resources has been demonstrated.