In this paper, we introduce the multi-stage configuration for the interpretation of user's queries in our spoken dialogue system. In this configuration, a recovery mechanism is used to detect and recover the errors arising from speech recognition. To efficiently incorporate with this recovery mechanism, a recognition scheme that can provide a compact word graph is developed. The compact word graph is generated through a pruning method based on the N-best sentence score. Instead of setting threshold, we use the N-best sentence score to select word hypotheses of the compact word graph.